In 2025, the autonomous vehicle (AV) race accelerated sharply. Waymo’s weekly rides increased more than fivefold in just under a year, putting it neck and neck with China’s Apollo Go in cumulative autonomous miles. Meanwhile, Tesla launched its robotaxi pilot mid-year and is currently gearing up for mass production of its Cybercab, targeting 2 million units annually. 

In our previous note, Driving Towards Autonomy (2024), we provided a primer on the autonomous vehicle space. This report builds on that foundation, examining key developments over the past year. We explore the surge in autonomous ride-hailing, expanding partnerships, unit economics and why we believe the AV sector is now at an inflection point.

Autonomous ride-hailing surge

Ride-hailing has seen strong momentum over the last 12 months. Waymo’s cumulative autonomous miles reached 100 million in July 2025 (Figure A), supported by more than 250,000 trips per week across five major US cities and a fleet of over 1,500 vehicles.

Figure A. Waymo’s cumulative autonomous miles 

Source: Waymo, X

Waymo has also taken market share from other ride hailing services, surpassing Lyft in San Francisco in late 2024 (Figure B). Waymo now plans to add an additional 2,000 Jaguar I-PACE vehicles in 2026 and expand its services to more cities. Additionally, their new integration plant in Mesa, announced in May 2025, is expected to be capable of building tens of thousands of fully autonomous Waymo vehicles per year when fully operational.

Figure B. Waymo market share versus competitors in San Francisco

Source: Bond, Trends – Artificial Intelligence 2025

Chinese autonomous mobility companies have also made strong progress, led by Baidu’s Apollo Go, often described as the Waymo of China. As of May 2025, Apollo Go had completed 11 million cumulative ride-hailing trips, slightly ahead of Waymo’s 10 million at the same time (Figure C). In Q2 2025 alone it delivered 2.2 million rides, representing 148% year-on-year growth. Since February 2025, Apollo Go has been operating fully driverless across China with no safety drivers and now serves 16 cities globally.

Figure C. Ride-Hailing Trips: Apollo Go versus Waymo

Sources: Waymo, Baidu, AlphaTarget

Technology, safety and scaling

Two main architectures have emerged in the autonomous race. Waymo, along with most peers, have adopted a multi-sensor fusion approach, typically combining cameras, LiDAR and radar with high-definition (HD) maps built in advance for each city. These detailed maps (lanes, signs, kerbs, crossings), which can provide centimeter-level accuracy, enable the vehicle to  localise precisely. Typically, these services run in geo-fenced areas with precise mapping and good network coverage, with the latter facilitating remote tele-operation as required. 

Tesla, by contrast, uses primarily cameras with an end-to-end neural-network approach and does not use HD maps for localisation. Note that both approaches use maps for navigation and routing, but Waymo and similar architectures depend on pre-built HD maps to precisely localise the vehicle and validate its sensor-based perception, while Tesla’s system determines position and driving decisions directly from camera feeds, supplemented by a microphone for detecting emergency sirens.

Waymo’s approach offers strong redundancy and precise localisation, which underpins proven commercial deployments. Waymo has also published data showing its vehicles are safer than the average human driver across several key metrics in the areas where it operates, with ‘91% fewer serious injury or worse crashes’ (Figure D). According to a TIME article in June 2025, Waymo’s Director of Field Safety, Matthew Schwall, believes the company’s autonomous vehicles are technically capable of operating safely in any American town, but widespread rollout depends on building local infrastructure and public trust.

Figure D. Waymo safety data

Source: Waymo

However, Waymo’s vehicles are still far from perfect drivers. There have been numerous reports of unpredictable behaviour, such as unexpectedly stopping in traffic. A Reuters article in August 2025 highlighted a Waymo driving into a flood and the passenger had to find a way out. Police said if the person had died, it could have led to a “serious criminal incident”. This highlights that the technology still needs refinement and has not quite yet reached the performance levels needed for widespread deployment. Additionally, the need to map out each new territory in incredible detail before autonomous operations has raised questions about its ability to scale at a meaningful pace. That said, Waymo is making progress on this front, with their 6^th generation vehicle “on track to begin operating without a human behind the wheel in about half the time”.

“Once we can make it basically work in a few cities in America, we can make it work anywhere in America. Once we can make it work in a few cities in China, we can make it work anywhere in China.”

Elon Musk, Tesla CEO (Q1 2025)

One of the benefits of Tesla’s approach is that fewer sensor types simplify hardware and lower costs. The approach is also more generalisable because it does not require detailed mapping before it expands to a new city. Tesla has taken the route of rolling out its autonomous technology as an Advanced Driver Assistance System. Thus, Tesla’s customers are able to use its Full Self Driving (Supervised) solution and report back on interventions, which allows the company to iteratively improve its generalised autonomy solution. Tesla thus also has orders of magnitude more data than other autonomous mobility players given its large existing fleet, allowing it to capture a long-tail of scenarios. Its vertical integration also means it can iterate and optimise faster.

“LiDAR is a fool’s errand, and anyone relying on LiDAR is doomed”

Elon Musk, Tesla CEO (April 2019)

Many, however, question whether a camera-only system can meet the safety threshold  for driverless operations. Musk argues that a multi-sensor approach increases complexity, potentially reducing signal-to-noise ratio (Figure E). Tesla’s “photons to control” approach utilises raw camera feeds, with their high dynamic range providing rich visual information, for both model training and inference at run-time. 

Figure E. Elon Musk’s posts on autonomous technology

Source: X

In June 2025, Tesla soft-launched its robotaxi service in Austin, with many riders reporting that the vehicles felt smoother than Waymos (e.g. fewer instances of phantom braking). However, the rollout is at a very small scale and by invitation only, so it is likely that participants were to some extent biased. Additionally, there were safety monitors in the passenger seat, though Musk has said these will be removed by the end of the year.

Musk has also outlined exceptionally ambitious scaling targets. On the Tesla Q2 2025 earnings call, he stated that he expects the number of autonomous vehicles in operation to grow at a hyper-exponential rate and projected that Tesla would offer autonomous ride-hailing to half the US population by the end of the year. He has also previously said that its new Cybercab vehicle, designed for ride-hailing, is targeted for volume production in 2026, targeting  2 million units per year and potentially 4 million at full scale. These targets illustrate the scale of Tesla’s ambition relative to its peers, who generally speak in terms of thousands of units rather than millions.

Currently, the jury is still out on which approach will ultimately prove superior. Tesla must demonstrate it can safely operate its robotaxis without safety monitors and show it can scale meaningfully, even if not fully to its most ambitious targets. Meanwhile, Waymo will need to refine its technology to reduce occasional unpredictable behaviour and improve the efficiency of scaling operations over time. Both companies face critical milestones and the path to broad autonomous adoption will hinge on how well each can meet these challenges.

Expansion via partnerships

“Leveraging our partners’ local market presence, we can accelerate market entry across different continents and achieve faster deployment while maintaining a cost-efficient asset-light business model.”

Robin Yanhong Li, Baidu CEO (Q2 2025)

To accelerate deployment, autonomous driving companies are increasingly forming partnerships with established ride-hailing services and automotive players. In the US, Waymo continues to collaborate with both Uber and Lyft across multiple cities, integrating its autonomous vehicles directly into their ride-hailing apps. In Japan, Waymo announced a partnership in late 2024 with taxi platform GO and operator Nihon Kotsu, where vehicles will initially be manually driven in Tokyo to refine local performance before autonomous operations begin. Waymo has also expanded its OEM relationships, including with Hyundai, whose IONIQ 5 SUV will join the Waymo One fleet and Toyota, with which it is co-developing technologies to enhance both personally owned vehicles and future fleet models.

Similarly, Apollo GO has also struck a range of partnerships to further expand its operations. For instance, in August 2025 it partnered with Lyft to deploy autonomous rides across Europe, with initial deployments planned for Germany and the UK in 2026, pending regulatory approval. The agreement sees the fleet scaling to thousands of vehicles in the following years, with Apollo Go providing the vehicles and comprehensive technical support and Lyft owning the operational value chain. It also struck a similar deal with Uber in July 2025 for a range of markets outside of the US and mainland China. Initial target markets are Asia and the Middle East, with Dubai emerging as a key hub given its goal for at least 25% of all trips to be driverless by 2030.

These partnerships with ride-hailing platforms allow autonomous driving companies to focus on their core expertise while leveraging partners’ established user bases and operational capabilities. This creates natural synergies and removes the need for autonomous mobility players to spend heavily and compete on customer acquisition. Going forward, we would expect these types of partnerships to continue and become a standard go-to-market channel.

Chinese autonomous players are also increasingly looking to shift to an asset-light approach, choosing not to own the vehicles themselves. This model reduces capital intensity and transfers fleet management to partners better equipped to maximise utilisation and minimise operating costs, which should ultimately enable more efficient scaling.

The road to profitability

“Looking into the longer-term, we see a clear path to profitability as hardware and labour costs keep coming down and our growing operational scale brings more efficiencies.”

Robin Yanhong Li, Baidu CEO (Q1 2025)

A key question for the autonomous vehicle industry is if and when robotaxi operations can achieve profitability. While removing the human driver eliminates a major cost, the high expense of autonomous hardware and ongoing operating overheads, including the need for tele-operators, have historically made positive unit economics difficult to achieve.

A Wall Street Journal analysis in December 2024 estimated the cost of a Waymo vehicle at roughly US$125,000, comprising about US$50,000 for the Jaguar I-PACE base car and the remainder for autonomous components, including LiDAR, cameras, radar and external audio receivers (EARs). Assuming US$15–20 billion in capital investment and a fleet of 20,000 vehicles, the analysis suggested a breakeven period of around 10 years.

“With 13 cameras, 4 LiDAR, 6 radar and an array of external audio receivers (EARs), our new sensor suite is optimised for greater performance at a significantly reduced cost, without compromising safety.”

Satish Jeyachandran, Vice President of Engineering Waymo (August 2024)

Although Waymo faces less pressure to cut costs given its parent Alphabet has significant resources, cost reductions have been an important focus. When it announced its 6^th generation Waymo Driver in August 2024, cost optimisation was a key highlight. Additionally, its partnership with Hyundai, adding the more affordable IONIQ 5 relative to the Jaguar I-PACE, suggests a focus on improving unit economics.

“RT6 is now running at a meaningful scale across multiple cities. And its unit cost is below $30,000, far better than anyone else on the planet.”

Robin Yanhong Li, Baidu CEO (Q1 2025)

In contrast, Baidu’s Apollo Go has achieved meaningfully lower vehicle costs and already reached positive unit economics in China, despite significantly lower average fares. In our discussions with another Chinese autonomous player, they have also seen significant cost reductions in recent times on multiple fronts. For instance, the hardware has become much more affordable, especially since they were able to move from expensive mechanical LiDAR to the more affordable solid-state and semi-solid state LiDAR. Additionally, as the AV industry continues to expand, the supply chain has benefitted from economies of scale, driving down costs. Operating expenses have also fallen through more efficient algorithms that consume less power and by reducing remote safety supervision to one operator per 30 vehicles.

This provider estimated total daily costs of around RMB300 (US$42) for its next generation vehicle, including OPEX and depreciation (Figure F). With taxi drivers in Chinese markets earning up to RMB600 (US$82) per day and autonomous fleets capable of continuous operation, they too appear well positioned to achieve positive unit economics. As fleets scale, fixed overheads should become an even smaller share of total costs, paving the way for profitability at the organisational level. 

Figure F. Autonomous vehicle unit cost for a Chinese autonomous mobility company

Source: Chinese autonomous mobility company

Tesla’s Cybercab is also set to be highly cost effective. Musk has stated the cost was expected to be below $30k, with no pedals, steering wheel or driver dashboard. It is designed for a “gentle ride”, with lower top end speed and more efficient tires than its other models, optimised fully for cost per mile, estimated at around $0.3-$0.4.

While vehicle costs are falling, early market data shows strong willingness to pay. A study by Obi found that Waymo rides were priced roughly 30–40% higher than competing services (Figure G). Among surveyed riders who had taken a Waymo, 70.2% said they preferred it, 16.7% were indifferent and 13.2% favoured traditional ride-hailing apps.  

Figure G. Price comparison between Waymo and other ride-hailing apps

Source: Obi, The Roach Ahead: Pricing Insights On Waymo, Uber and Lyft
Notes: n: 88,998, Dates: 25 March – 25 April 2025, Place: San Francisco

When respondents were asked why they preferred the Waymo, 52% of respondents said it was because they did not need to speak to the driver (Figure H). It is also likely that some riders feel safer without a human driver, though that was not an option on the questionnaire. We assume these factors can to some extent explain its growing market share and its ability to command a premium.

Figure H. Survey for why riders preferred Waymo

Source: Obi, The Roach Ahead: Pricing Insights On Waymo, Uber and Lyft

“We are moving towards a future where taking a robotaxi will be half the cost of taking a taxi today.”

Robin Yanhong Li, Baidu CEO (July 2022)

Although riders are currently willing to pay a premium, we do not expect elevated pricing to persist. In the long run, we would expect fares to fall materially below current ride-hailing rates as cost savings from removing human labour and using purpose-built vehicles are competed away. However, lower prices should drive significantly higher adoption and utilisation rates, as autonomous services become more accessible to a broader segment of consumers.

We still expect healthy margins over time, however. We do not anticipate the autonomous driving market to resemble today’s car industry with over 100 brands and structurally low profitability. Once leading players achieve scale, accumulate large amounts of proprietary driving data and demonstrate superior safety performance, barriers to entry should rise sharply, limiting competition and supporting healthy margins. At the same time, we do not foresee monopoly-like outcomes, as large national governments are unlikely to allow a single foreign operator to dominate domestic mobility infrastructure and are expected to support local champions.

Sizing the opportunity

Autonomous vehicles will potentially impact multiple trillion-dollar markets including transportation, logistics and urban infrastructure. This potential stems from fundamental changes to cost structures and asset utilisation across these sectors, although the early stage of the technology makes precise market sizing challenging.

Since autonomous technologies are still in the very early stages of their rollout, the estimates for total addressable market (TAM) vary significantly across forecasts. For instance, according to Fortune Business Insights, the global autonomous vehicle market was valued at just over US$1.9 trillion in 2023 and is expected to grow to US$13.6 trillion by 2030  (representing a CAGR of 32.3%)! On the other end of the spectrum, Precedence Research has a relatively conservative estimate for the Autonomous Vehicle Market at US$158.3 billion in 2023, set to grow to US$2.75 trillion by 2033 (Figure I).

Figure I: Autonomous Vehicle Market Size, 2023 to 2033

Source: Precedence Research

Conclusion

The autonomous vehicle (AV) market remains nascent, with leading robotaxi fleets currently numbering around 1,000–1,500 vehicles and most new partnerships involving only a few thousand units. This is orders of magnitude smaller than the global vehicle total, estimated at 1.6 billion. However, growth is accelerating as technology advances, regulatory support increases and partnerships proliferate across regions. Crucially, we see emerging positive unit economics as a catalyst, giving companies the confidence and economic capacity to scale more aggressively. We therefore believe the industry is now at an inflection point and we would not be surprised to see fleet sizes doubling or tripling annually for many years to come.

At AlphaTarget, we invest our capital in the most promising disruptive businesses at the forefront of secular trends; and utilise stage analysis and other technical tools to continuously monitor our holdings and manage our investment portfolio. AlphaTarget produces cutting-edge research and our subscribers gain exclusive access to information such as the holdings in our investment portfolio, our in-depth fundamental and technical analysis of each company, our portfolio management moves and details of our proprietary systematic trend following hedging strategy to reduce portfolio drawdowns.  To learn more about our research service, please visit alphatarget.com/subscriptions/.

“[T]he larger driver of our everyday activity is the fact that the software industry is completely changing with AI. It feels like the AI era of the past two years just makes the SaaS era feel like sleeping days.” 

Ivan Zhao, Notion CEO (11 August 2025, The Verge)

“It’s been eight of the most exciting months, I think, of my career.”

Marc Benioff, Salesforce CEO (29 August 2025, The Logan Bartlett Show)

An agentic software future

The current artificial intelligence (AI) revolution has ignited widespread debate about its transformative potential across industries, with software playing a central and pivotal role in these changes. Many observers express concerns that AI could disrupt the very foundations of software as we know it, rendering the current software paradigm obsolete. This fear stems from the belief that AI’s generative capabilities could automate code generation to the point where bespoke applications are created with ease, bypassing the need for ongoing subscriptions to SaaS platforms. Additionally, some industry experts foresee a future where traditional software becomes redundant, as instead of interacting with a range of siloed software applications, users will interact with an agentic interface through which they can express their intent and have agents work in the background to obtain the desired outcome.

“I think the notion that business applications exist, that’s probably where they’ll all collapse, right, in the agent era.”

Satya Nadella, Microsoft CEO (12 December 2024, BG2)

High-profile voices have amplified these fears. In a recent social media post (Figure A), Musk suggested that phones and computers would effectively become edge nodes for AI, directly rendering pixels without the need for a traditional operating system or conventional apps. In that same exchange, he agreed with Replit’s CEO, Amjad Masad, who commented that AI could generate apps on demand, with agents eating traditional apps. Replit itself exemplifies this shift: its vibe coding (i.e., “rapid prototyping”) tool allows users to build and deploy apps in a fraction of the time compared to traditional software development, collapsing traditional coding barriers and enabling anyone to be a software developer.

Figure A: Elon Musk’s post on X

Source: X

We believe the AI revolution will significantly impact existing software, but that does not necessarily mean the demise of established software vendors. Presently, vibe coding tools excel at code generation and prototyping, yet have fallen short in delivering fully polished, production-ready products. We suspect that time will come too, but even then, incumbents possess moats that cannot easily be overcome in the short run. These include network effects, deeply embedded workflows and troves of user data. For instance, CRM is not just a tool but a living ecosystem of customer relationships and integrations that vibe coding cannot simply replicate. However, existing software will have to evolve significantly to meet user expectations if it is to stay relevant in the long run.  

In an August 2025 report, Gartner forecasts the future evolution of AI in enterprise software, as shown below (Figure B). Although timelines are highly uncertain, we think this forecast provides a good outline of the many steps (and hurdles) and time taken before we get to a world where agentic software is mainstream.

·     Stage 1. By the end of 2025, the majority of enterprise apps will have embedded AI assistants to simplify tasks and enhance user productivity. AI assistants are generally human triggered.

·       Stage 2. By 2026, 40% of enterprise apps will incorporate task-specific AI agents capable of handling complex, end-to-end processes independently, such as automating cybersecurity responses or other specialised functions.

·     Stage 3. By 2027, one-third of agentic AI implementations will have different agents collaborating within individual applications, combining diverse skills to manage intricate tasks.

·    Stage 4. By 2028, there will be AI agent ecosystems that collaborate across multiple apps and business functions. One-third of user interactions will occur through agentic front ends.

·         Stage 5. By 2029, at least 50% of knowledge workers will be able to create AI agents on demand.

Figure B: Gartner’s forecast for agentic AI in enterprise apps

Source: Gartner (August 2025)

An example of this agentic evolution is the introduction of Salesforce’s Agentforce, which it launched in September 2024. This AI-powered autonomous agent platform connects to enterprise data and handles tasks across sales, service, marketing and commerce. Agentforce has seen rapid adoption, with over 12,500 deals closed by July 2025, of which 6,000 were paid. Below is an illustration of its enterprise stack (Figure C), composed of a data layer (Data Cloud), a core application layer (Customer 360) and the agentic layer (Agentforce) which orchestrates work on top. We think this is a good illustration of where the enterprise stack is heading, though we acknowledge that there will likely be variations across different companies and industries. A key challenge for incumbent companies like Salesforce however will be adapting these innovations around legacy systems rather than building AI-native solutions from scratch.

Figure C: Salesforce’s Agentforce

Source: Salesforce

The fruits of thy labour

“[T]he overall addressable market went from software to software plus labour, which is 10 times bigger.”

Yamini Rangan, HubSpot CEO (5 May 2025, Grit)

It is important to note that the rise of AI is not just a challenge for software vendors to overcome in order to stay relevant, but also a significant opportunity. Those who innovate early and integrate AI capabilities into their products can offer more advanced services, effectively monetising the additional value they create. In other words, the leaders in this AI-driven evolution can potentially charge for the extra “intelligent” work their software does, creating significant new revenue streams.

At the same time, there is an open question about how much of this value will be captured by the software vendors relative to the infrastructure and model providers e.g. Microsoft and OpenAI, who supply the inference capabilities. Although inference is getting cheaper, with some estimating it is becoming 10x cheaper per year, the amount of inference needed per task is also growing extremely quickly, especially as models become more advanced and do more work. In an article by The Wall Street Journal, they provided an estimate of tokens needed per task below, showing how inference requirements increase drastically with task complexity.  

• Basic chatbot Q&A: 50 to 500 tokens

• Short document summary: 200 to 6,000 tokens

• Basic code assistance: 500 to 2,000 tokens

• Writing complex code: 20,000 to 100,000+ tokens

• Legal document analysis: 75,000 to 250,000+ tokens

• Multi-step agent workflow: 100,000 to one million+ tokens

The article also commented on how this is impacting margins: Ivan Zhao, CEO of productivity software company Notion, said his business had a gross margin of 90% two years ago, fairly standard for SaaS companies. However, approximately 10 percentage points of that margin now goes to the AI companies that power its latest offerings.

It is difficult to forecast the long-term margin impact of inference on SaaS companies, as it will depend on several factors. These include the market dominance of hardware and model providers, which determine their pricing power. Another determinant of inference costs will be the extent to which hardware and software efficiencies improve relative to inference demand. However, if the history of traditional cloud computing by big tech is any guide, inference providers are likely to capture a healthy share of the economics.

Still, we think innovative software vendors will see more upside than downside. They will be able to offer significantly more value to their clients by automating work and reducing labour costs, value that can be monetised at a strong premium. Additionally, as agent-driven systems become embedded in everyday workflows, almost like training an employee who learns a company’s specific needs, vendors could become much harder to dislodge, strengthening their moats. This, in turn, should allow them to raise prices and improve margins. ServiceNow, for instance, exemplifies this opportunity, with the company’s NowAssist AI products commanding 20%+ price premiums over standard offerings, while driving significant revenue growth.

Monetisation revamp

As software increasingly performs work through AI inference, software companies will also need to restructure their monetisation models. SaaS companies will need to shift away from their reliance on seat-based revenue towards usage-based revenue to align monetisation with value delivered. However, this transition is fraught with challenges. Seat-based pricing is more straightforward, scaling with user count. Usage-based models, however, introduce unpredictability, as costs fluctuate with consumption, making forecasting difficult for customers and businesses. It requires real-time systems to track potentially unbounded spend, unlike monthly batch processes. Complex discount structures for enterprise contracts add further intricacy, and data must be stored accurately and accessibly to maintain financial integrity and future pricing flexibility.

Operationally, incentive structures will transform. Sales teams must prioritise high-usage clients over seat counts, necessitating new compensation models. Product teams need to focus on unlocking use cases that drive usage, requiring agility to adapt to evolving customer needs. This transformation demands top-down leadership from the CEO to align sales, product, and finance teams. The right pricing model is still being figured out, with no universal solution yet, and we expect this will take time as AI models and hardware continue to rapidly improve. As seen with Salesforce’s Agentforce, it has reportedly changed its pricing model three times in a single year, reflecting rapid experimentation to find the optimal approach.

Despite challenges, usage-based pricing offers immense opportunity. By aligning monetisation with usage, companies achieve better product-market fit, capturing more value as customers derive greater benefit, fostering a flywheel of growth and innovation.

The agentic organisations

“I am on a mission to make Salesforce an agentic enterprise.”

Marc Benioff, Salesforce CEO (29 August 2025, The Logan Bartlett Show)

In addition to evolving their products and monetisation models, software vendors will also need to rethink their organisational structures. As Marc Benioff from Salesforce highlighted, it’s not just about making the product agent-driven, but about making the entire organisation agent-driven. In a recent podcast, he mentioned reducing Salesforce’s support staff from around 9,000 to 5,000 thanks to AI agents, reducing costs substantially.  

Likewise with sales, AI agents are reshaping how leads are managed and converted. Benioff explained that Salesforce had accumulated over 100 million uncontacted leads over 26 years due to insufficient human resources. However, now agents call back every person who calls them, handling more than 10,000 leads weekly, engaging in conversations and funnelling them into their pipeline. This highlights how agents are not just helping cut costs, but also helping grow revenues more efficiently. We imagine there will be many more use cases where agents will enhance efficiencies significantly. We therefore see significant upside potential for companies that can transform themselves to become more agent driven.

Sizing the opportunity

Cutting-edge software firm Palantir estimates it will be able to grow its revenue by 10x over the next 5 years, while simultaneously reducing headcount. This is a testament to the potential efficiency gains of AI. Although most firms will likely not be as operationally efficient as Palantir, we believe the potential for software to automate services and displace human labour will still lead to a substantial TAM expansion. According to Battery Ventures’ 2024 State of the OpenCloud report, the AI software TAM is worth approximatley US$4 trillion (Figure D).

Figure D: The AI Software Market Opportunity

Source: Battery Ventures, State of the OpenCloud Nov 2024

Conclusion

The current AI paradigm poses one of the greatest challenges the software industry has faced, forcing vendors to rethink many aspects of their business. Yet it also represents a powerful accelerant. Vendors could see substantial revenue uplifts by monetising the intelligent work their software performs, while also capturing significant operational efficiencies within their own organisations. Existing moats give incumbents valuable time to adapt, and as software evolves from being a critical tool to performing critical workflows, these moats may in fact deepen, making vendors harder to displace. However, Darwinism will take effect. The winners will be those agile enough to adapt quickly and embrace AI holistically, not just at the product level, but across their entire organisations.

In our view, companies that are already geared towards usage-based pricing models, will naturally benefit in this evolving AI landscape. We believe usage will rise as AI performs more tasks and we are already seeing signs of strong growth for software infrastructure companies, where usage based revenue models are native. Seat-based models on the other hand may face headwinds as businesses become more efficient and may as a consequence see less seat growth. We are therefore more cautious on vendors that remain seat-based without a clear path to transition toward usage-based models, particularly where their products have limited potential to perform meaningful work that could be monetised.

Additionally, we have a preference for infrastructure software businesses with deeply embedded workflows. Such companies are less easily disrupted than application software vendors and have more breathing room to evolve without immediate competitive threats. We also see cybersecurity as a particularly promising area. As companies expand their AI capabilities, their attack surfaces and vulnerabilities will grow. This will create more demand for advanced cybersecurity solutions. With adversaries also leveraging AI, cybersecurity firms that can help clients secure these new environments are well-positioned to benefit.

At AlphaTarget, we invest our capital in some of the most promising disruptive businesses at the forefront of secular trends; and utilise stage analysis and other technical tools to continuously monitor our holdings and manage our investment portfolio. AlphaTarget produces cutting-edge research and our subscribers gain exclusive access to information such as the holdings in our investment portfolio, our in-depth fundamental and technical analysis of each company, our portfolio management moves and details of our proprietary systematic trend following hedging strategy to reduce portfolio drawdowns.  To learn more about our research service, please visit alphatarget.com/subscriptions/. 

Quantum computing is emerging as the next frontier in technology and it has the potential to solve problems that today’s computers cannot touch; transforming industries from cryptography to drug discovery. Yet, despite billions invested and headlines touting breakthroughs, practical quantum advantage remains years away. Investors should watch this space carefully but resist the temptation to chase the hype.

This article breaks down the key components of the quantum computing ecosystem, from processors and infrastructure to software and real-world use cases, exploring how investors and business leaders can participate in its growth. We cover the main hardware technologies, supporting infrastructure providers, early commercial applications, current public investment opportunities, risks to watch and some principles for investing in the space. 

Understanding quantum computing

Quantum computing is an emerging field in computer science and engineering that leverages the principles of quantum mechanics to tackle problems that exceed the capabilities of even the most powerful classical computers. The field of quantum computing includes a range of disciplines, including quantum hardware and quantum algorithms.

Classical computers, from smartphones to supercomputers, process information using bits (i.e., they are in one of the two states ‘1’ or ‘0’). These computers handle tasks by manipulating bits through logical operations, often in parallel across multiple processors to speed up computations. However, their performance is limited by the need to evaluate solutions step-by-step for complex problems, such as factoring large numbers or simulating chemical reactions.

Quantum computers use quantum bits (qubits), which differ fundamentally from classical bits. Due to a property called superposition, a qubit can represent 0, 1, or a combination of both states simultaneously. This enables quantum computers to process multiple potential solutions at once. Additionally, qubits can be entangled, meaning the state of one qubit is linked to another, allowing coordinated computations that classical systems cannot replicate. These properties make quantum computers potentially far faster for certain tasks such as cryptography, molecular modelling or risk analysis, though they are not universally superior and require specialised algorithms to outperform classical systems.

Google’s Quantum AI Lab

Source: Google

Hardware: The core engine

As discussed above, qubits are at the heart of the quantum computing revolution. Unlike bits that are binary, qubits operate using quantum superposition and entanglement, allowing them to represent multiple states simultaneously. A key challenge is that qubits are extraordinarily fragile; the slightest vibration, temperature change, or electromagnetic interference causes “decoherence” (i.e., the qubit loses its quantum properties).

Currently, three main approaches compete to build stable qubits:

• Superconducting qubits (used by IBM, Google, Rigetti): Rely on superconducting circuits at millikelvin (extremely cold) temperatures. Benefits include fast gate speeds and integration with CMOS-compatible fabrication, but they suffer from decoherence (losing quantum properties within microseconds) and scaling issues.
• Trapped ion qubits (IonQ, Quantinuum): Leverage ions suspended in electromagnetic fields. Highly accurate with longer coherence times, but gate speeds are slower and hardware is more difficult to scale.
• Photonic qubits (PsiQuantum, Xanadu): Use photons as carriers of quantum information, promising scalability and room-temperature operation, though achieving reliable operation remains a challenge.

The modalities have different commercial prospects and timelines to scale. Investors should view this like the early days of semiconductors, with multiple competing technologies and no clear winner yet. Rather than focus on one approach, a prudent approach might be to consider companies with strong IP portfolios and proprietary fabrication capabilities. Partnerships with national labs or hyperscalers (e.g., Microsoft + Quantinuum) signal credibility. Furthermore, as discussed in the next section, the picks and shovels of this area might also be worth further investigation. 

More than just chips

Quantum computing systems are extremely delicate, and their operation requires a complete redesign of the computing stack. That includes:

• Cryogenic refrigeration: Dilution refrigerators are essential for superconducting qubits, often operating below 15 millikelvin. Bluefors and Oxford Instruments dominate this niche, offering high-margin components with little competition.
• Control and readout electronics: Precise microwave pulses must control each qubit, which requires specialised signal generators and converters. Keysight and Zurich Instruments provide these critical components. 
• Quantum interconnects: Low-loss coaxial cables, optical fibre links for photonic setups, and waveguides are all part of the infrastructure.
• Quantum-safe networking: Emerging players are enabling quantum key distribution (QKD) and post-quantum cryptographic protocols to secure classical systems against quantum decryption threats.
• Quantum cloud platforms: Big tech now offers quantum hardware via the cloud. IBM Quantum, Amazon Braket, and Azure Quantum act as intermediaries for enterprises to experiment, without capital-intensive ownership. This “quantum-as-a-service” model generates recurring revenue and lowers barriers to adoption.

Infrastructure suppliers (cryogenics, control hardware and packaging) are cash-flow positive and may see inorganic acquisition from hyperscalers or defence contractors. 

Emerging use cases

Quantum computing is poised to augment classical computing by addressing specific use cases more efficiently. 

In finance, current risk models can analyse thousands of scenarios. Quantum computers could analyse millions simultaneously, transforming option pricing, arbitrage strategies, risk analytics, and Monte Carlo simulations. For instance, JPMorgan Chase estimated that quantum computers could accelerate Monte Carlo simulations used in derivative pricing models by 1,000 times. Early pilots with banks show promising results.

In pharmaceuticals, quantum computing’s potential to model protein folding, ligand binding, and drug target interactions is expanding rapidly, with partnerships such as Roche collaborating with quantum computing firm Cambridge Quantum. Materials science benefits from quantum-powered molecular simulations that accelerate the discovery of new catalysts, with Chevron and BASF actively investing in this space.

The energy sector looks to quantum computing to optimise power grids, simulate fusion processes, and advance battery chemistry. These efforts are strongly supported by government initiatives focused on achieving net-zero carbon emissions.

Logistics applications include vehicle routing, warehouse optimisation, and scheduling, leveraging hybrid quantum-classical solvers that are gaining traction in the industry. 

Finally, cybersecurity is emerging as a critical area for quantum computing technology, focusing on post-quantum cryptography, quantum key distribution, and secure authentication to counteract future quantum decryption threats. This is driven by evolving NIST (National Institute of Standards and Technology) standards and increasing national security interests.

Firms with pilot projects or co-innovation partnerships in these domains are best positioned to capture early revenues, especially those nearing quantum computing advantage in their use cases.

Themes & opportunities

The quantum computing stack is broad, opening up multiple avenues of capital allocation:

• Quantum Computing Hardware Companies: Publicly listed companies include IonQ (IONQ), Rigetti (RGTI), and D-Wave (QBTS). In this space, one ought to monitor product roadmaps versus hype.
• Infrastructure and Tooling Suppliers: Cryogenics (Bluefors), microwave components (Keysight), dilution refrigerators (Janis Research), and control systems (Zurich Instruments). These firms are typically overlooked but benefit from early-stage adoption and defence applications.
• Quantum-as-a-Service (QaaS) Providers: Microsoft Azure Quantum, Amazon Braket, and IBM Cloud are developing pay-as-you-go pricing models. They offer recurring SaaS-style revenue with high customer switching costs.
• Quantum Computing Middleware & Software Tooling: SDKs like Qiskit (IBM), Cirq (Google), and tket (Quantinuum). Compilers, noise mitigation tools, and quantum optimisers are early bets with parallels to early AI tooling stacks.
• Application-Level Startups: While mostly private, those that go public or get acquired will benefit from early proven vertical quantum computing advantage.

The key players 

Quantum computing is still in its early innings as a commercial industry, but a handful of companies already exist in the public markets, either as pure plays or through broader technology portfolios. For investors, these companies represent a mix of long-term potential, near-term uncertainty and varying levels of exposure to quantum computing. 

IonQ (NYSE: IONQ)
IonQ was the first quantum computing company to go public via a SPAC in 2021 and remains one of the few true “pure plays.” It develops trapped-ion quantum processors, known for longer coherence times and high gate fidelity. IonQ offers hardware access via AWS, Azure, and Google Cloud. It is starting to generate revenue (US$43 million TTM) through quantum-as-a-service and research contracts. Its market capitalisation is US$9.8 billion, reflecting high expectations.

Rigetti Computing (Nasdaq: RGTI)
Rigetti develops superconducting quantum processors and hybrid classical-quantum systems. It builds its own hardware and cloud platform, aiming for vertical integration. Despite operational challenges, it has secured government contracts and research partnerships. TTM revenue is US$9 million and its market capitalisation is US$3.3 billion, reflecting high expectations.

D-Wave Quantum (NYSE: QBTS)
D-Wave focuses on quantum annealing, a specific type of quantum computing best suited to optimisation problems. Its technology is deployed commercially in logistics and manufacturing and is accessible via its Leap cloud platform. TTM revenue is US$21 million and its market capitalisation is US$4.9 billion, reflecting high expectations.

Arqit Quantum (Nasdaq: ARQQ)
Arqit focuses on quantum-safe encryption and security via its QuantumCloud platform. It operates at the intersection of quantum computing and enterprise security, appealing to investors interested in post-quantum cryptography. Revenue is still tiny and the market capitalisation is US$545 million.

Honeywell (Nasdaq: HON)
Though not a pure quantum computing play, Honeywell owns Quantinuum (from Honeywell Quantum Solutions and Cambridge Quantum). Quantinuum develops trapped-ion systems and middleware/encryption tools, offering indirect quantum computing exposure.

IBM (NYSE: IBM)
IBM is a pioneer in quantum computing R&D, offering superconducting qubit systems via IBM Quantum and an open-source software development kit (Qiskit). Its roadmap targets scaling to thousands of qubits and integrating quantum computing with cloud services.

Alphabet (Nasdaq: GOOGL)
Google’s Quantum AI group has made landmark achievements, including controversial but significant quantum computing supremacy claims. For instance, the company’s 105-qubit Willow chip (unveiled in December 2024) is said to perform a random circuit sampling task in under five minutes, which Alphabet estimates would take a supercomputer 10 septillion (10^25) years. Critics argue the task lacks practical use and classical algorithms may close the gap. While experimental, Google’s quantum computing efforts are part of its broader innovation portfolio.

NVIDIA (Nasdaq: NVDA)
Not a quantum computing hardware company but an enabler through GPU-based quantum simulators and hybrid classical-quantum infrastructure like cuQuantum. NVIDIA represents a “picks and shovels” investment in the quantum computing ecosystem.

Challenges & risks

Even though quantum computing appears to have a long growth runway, investors must approach this area with realistic expectations:

• Scalability limits: Many architectures remain under 100 qubits; fault-tolerant systems require thousands.
• Error correction overhead: Surface code error correction demands 1,000+ physical qubits per logical qubit.
• Hardware fragility: Qubits are sensitive to noise, vibrations, and temperature fluctuations.
• IP uncertainty: Overlapping patents and licensing disputes may slow innovation.
• Geopolitical tension: Quantum computing is a national security priority for the US, China, EU, and UK. Export controls and regulations could impact deals.
• Time-to-profitability: Many firms may take 5–10 years before they generate meaningful cash flows and profits.

In our view, in many ways quantum computing is similar to biotechnology, with potentially highly binary outcomes and long timelines.

Sizing the opportunity

Beyond the technology and individual companies, the potential growth of the quantum computing opportunity demands attention. As shown in Figure A, Precedence Research projects that the quantum computing market will reach US$16.22 billion by 2034 from US$1.44 billion today, representing a compound annual growth rate (CAGR) of 31%. North America accounted for US$1.1 billion in 2024, and is expected to expand at a CAGR of 31% during the forecast period.

Figure A: Quantum Computing – long growth runway

Source: Precedence Research

Over the longer term, McKinsey & Company estimates that the quantum computing market could be worth US$45 billion to US$131 billion by 2040.

This opportunity spans multiple sectors. In pharmaceuticals, quantum modelling may significantly accelerate drug discovery and personalised medicine, unlocking billions in R&D productivity. In finance, faster and more accurate risk modelling, arbitrage, and portfolio optimisation are poised for major transformation. In materials science and energy, the simulation of atomic structures and chemical reactions could lead to breakthroughs in battery technology, sustainable fuels, and next-generation industrial catalysts.

Adjacent markets like quantum-safe cybersecurity, cloud-delivered QaaS platforms, and hybrid high-performance computing infrastructure further broaden the potential. In addition, government investment, including multibillion-dollar national quantum computing initiatives across the U.S., China, EU, and others, provides a baseline of funding that de-risks early-stage development.

For investors, this means that quantum computing could turn out to be a platform shift on par with classical computing, cloud and AI; opening the door to high returns for those who time the inflection point correctly.

Conclusion

Quantum computing stands at the threshold of reshaping global industries, from finance and pharmaceuticals to energy and logistics. The real economic opportunity lies not just in building quantum computers, but in constructing the infrastructure and software that enable scale. The best investments will combine long-term vision with near-term validation, and reward those who position capital accordingly.

Key considerations for investors include recognising that infrastructure returns often come early; while hardware may take years to generate meaningful ROI, companies providing essential tools, cryogenics, and quantum-as-a-service platforms are already producing cash flow. Valuing quantum computing companies requires a hybrid approach, measuring technical milestones such as qubit count and fidelity, intellectual property strength, customer pilots, and partnership depth alongside traditional fundamental metrics. Mid-layer enablers like refrigeration, signal control, and middleware SDKs appear to be underpriced, offering “picks and shovels” opportunities with less hype but stronger customer entrenchment. 

This market is akin to investing in AI infrastructure circa 2015, as it is still in the pre-explosion phase. Just as Nvidia and TSMC capitalised on the AI boom by supplying foundational infrastructure, quantum computing infrastructure players may also drive the next wave of computational transformation.

At AlphaTarget, we invest our capital in some of the most promising disruptive businesses at the forefront of secular trends; and utilise stage analysis and other technical tools to continuously monitor our holdings and manage our investment portfolio. AlphaTarget produces cutting-edge research and our subscribers gain exclusive access to information such as the holdings in our investment portfolio, our in-depth fundamental and technical analysis of each company, our portfolio management moves and details of our proprietary systematic trend following hedging strategy to reduce portfolio drawdowns.  To learn more about our research service, please visit alphatarget.com/subscriptions/.

Artificial intelligence (AI) is reshaping industries – from healthcare diagnostics to financial trading – but its transformative power depends on an often-overlooked foundation: AI infrastructure. This specialised ecosystem of hardware, software, and connectivity enables technologies such as chatbots, recommendation systems, and autonomous vehicles to operate effectively.

Unlike traditional cloud computing, which supports general applications such as e-commerce platforms or media streaming, AI infrastructure is purpose-built to meet immense computational demands of machine learning. Training sophisticated AI models, such as large language models (LLMs) that power virtual assistants, requires processing vast datasets across thousands of specialised processors. Deploying these models for real-time tasks, like generating instant recommendations, demands both speed and scalability. This dual challenge — training and deployment — necessitates unique architectural designs optimised for parallel processing, high-speed data transfer, and energy efficiency.

This article explores the core components of AI infrastructure, how it differs  from traditional cloud systems, the innovative companies driving its evolution, and the investment opportunities within this dynamic sector. From specialised chips to energy-efficient data centres, AI infrastructure is a critical enabler of technological progress. For investors, it offers a way to gain exposure to AI’s growth through the foundational technologies that will shape industries for decades.

Nuts and bolts of AI infrastructure 

AI infrastructure is the technological foundation for building, training, and deploying AI models, particularly large language models (LLMs) that fuel applications ranging from virtual assistants to personalised content recommendations. Unlike general-purpose computing, AI infrastructure is engineered for two distinct phases:

Training: The resource-intensive process of developing an AI model by processing vast datasets to optimise billions of parameters. This phase is comparable to a student spending months mastering complex subjects – requiring significant resources but occurring only once per model version.

Inference: The application of a trained model to deliver real-time outputs, such as answering user queries or generating recommendations. This phase prioritises speed, efficiency, and scalability – similar to a student instantly applying their knowledge across thousands of simultaneous tests.

Next, we examine some of the key components of AI infrastructure (Figure A). 

Figure A: AI Infrastructure Abstraction

AI accelerators: The computational engine

While traditional computing relies on Central Processing Units (CPUs), AI thrives on specialised accelerators designed for the parallel computations that dominate machine learning workloads.

• Graphics Processing Units (GPUs): Originally developed for graphics rendering, GPUs excel at performing thousands of calculations simultaneously. Nvidia currently leads the market with its Hopper and Blackwell architectures, designed for high-performance AI workloads. Its CUDA software ecosystem is widely adopted, but competitors like AMD (Instinct MI300) and Cerebras (Wafer-Scale Engine) are introducing alternatives that could challenge its position.
• Tensor Processing Units (TPUs): Google’s custom-designed AI chips, optimised for its TensorFlow and JAX frameworks. TPUs are application-specific integrated circuits (ASICs) engineered for machine learning, offering strong performance primarily within Google’s cloud ecosystem. 
• Emerging Alternatives: Companies like Cerebras (with its massive Wafer-Scale Engine), AMD (Instinct MI300, focussed on cost-competitive performance), and AWS (Trainium, optimised for cloud training) are challenging Nvidia’s dominance. While these alternatives show promise, it remains to be seen whether they can match Nvidia’s blend of performance and mature software ecosystem.

Storage and connectivity: The data lifeline

While AI accelerators drive computations, storage and connectivity ensure data moves swiftly and seamlessly to support machine learning workloads. These components are vital for training (analysing vast datasets) and inference (delivering instant outputs), serving as the lifeline that keeps AI systems operational.

• High-Performance Storage: AI models demand rapid access to enormous datasets. NVMe Solid State Drives (SSDs) enable swift data retrieval, minimising delays during training and inference. Distributed file systems, such as Lustre or WekaFS, facilitate parallel access to terabytes of data across multiple servers, supporting large-scale AI tasks.
• Advanced Connectivity: AI’s processors require constant, low-latency communication, particularly when thousands operate together. High-speed interconnects, like Nvidia’s InfiniBand, provide rapid data transfer between systems. Specialised switches from providers like Arista and Broadcom manage substantial traffic flows, ensuring smooth coordination across processors. PCIe technologies, essential within servers, enable high-speed connections between GPUs and other components, maintaining efficiency in dense computing environments.
• Emerging Innovations: Novel storage solutions, such as AI-optimised cloud-native systems from AWS and Google Cloud, combine speed and scalability. In connectivity, smart Network Interface Cards (NICs) from companies like Mellanox offload data tasks, improving performance. These advancements aim to rival established systems while addressing AI’s evolving needs.

Together, storage and connectivity form the backbone of AI infrastructure, ensuring data availability and processor synchronisation. Leaders like Pure Storage, Arista, Broadcom, and cloud providers are shaping this critical layer of the intelligence era.

Software ecosystem: The orchestration layer

Hardware alone isn’t sufficient, AI infrastructure requires sophisticated software:

• AI Frameworks: Leading tools like TensorFlow, PyTorch, and JAX provide programming environments for developing and optimising machine learning models.
• Orchestration Platforms: Kubernetes and similar cloud-native orchestration platforms distribute complex workloads across thousands of processors, ensuring scalability.
• MLOps Tools: Specialised tools like MLflow and Kubeflow manage the machine learning lifecycle, from development to deployment and monitoring.

Power and cooling: The physical foundation

The extraordinary computational demands of AI create unprecedented physical challenges:

• Energy Consumption: Large-scale training clusters can consume 10-20 megawatts of power, equivalent to powering approximately 10,000 homes.
• Advanced Cooling: Solutions like direct liquid cooling are essential, as air cooling alone cannot manage the heat generated by dense AI compute clusters.
• Sustainable Design: With rising energy costs and environmental concerns, innovations like renewable energy integration and power-efficient designs are becoming competitive differentiators for infrastructure providers.

The interplay between these components determines the performance and economics of AI infrastructure. For investors, understanding how these elements fit together reveals potential bottlenecks and opportunities across the AI stack.

The rise of specialised AI infrastructure players

The explosive demand for AI compute has driven the emergence of a new breed of technology providers focused exclusively on AI infrastructure. Unlike traditional cloud providers like AWS, Microsoft Azure, and Google Cloud, which serve diverse workloads from web hosting to enterprise databases, these specialised players build infrastructure optimised specifically to meet AI’s unique requirements.

Traditional cloud infrastructure, with its general-purpose CPUs and standard networking, was not designed for the intensive matrix calculations of neural networks or the massive data exchanges required by distributed training. While hyperscalers are rapidly expanding their AI offerings, specialised providers have gained momentum by focusing exclusively on optimising for AI workloads — offering superior performance, cost efficiency, and access to scarce compute resources.

A key factor driving this specialisation is the current scarcity of high-performance AI accelerators, particularly Nvidia’s coveted H100 and Blackwell GPUs. Access to these chips has become a strategic advantage, with the most successful providers securing multi-year allocations through partnerships with Nvidia. This supply constraint has created a market dynamic where specialised providers with secured GPU supply can command premium pricing and long-term contracts.

In this section, we review some of the key players in this emerging segment. 

CoreWeave

CoreWeave operates over 250,000 GPUs in 32 data centers, making it a top provider of AI infrastructure for companies like Microsoft and OpenAI. Its platform uses Nvidia’s fast networking and a system called SLURM on Kubernetes to manage workloads efficiently. This setup delivers up to 20% better performance (Model FLOPs Utilisation, or MFU) than competitors, accelerating AI training and processing.

CoreWeave’s business is built on long-term contracts, with a US$27 billion backlog (14 times its 2024 revenue of US$1.92 billion) reflecting both the supply-constrained market and customer confidence. These contracts typically span over four years, providing revenue visibility but introducing customer concentration risk, with two clients accounting for 77% of 2024 revenue.

The company’s recent acquisition of Weights & Biases enhances its software stack, offering developer tools for model training and positioning it as a full-stack provider rather than merely a hardware operator. CoreWeave has secured Nvidia GPU allocations through partnerships, supporting its global expansion. However, its reliance on Nvidia chips exposes it to supply chain risks, and competitors are exploring alternative accelerators.

Lambda Labs

While privately held Lambda Labs operates at a smaller scale than CoreWeave, it has carved out a niche serving startups, researchers, and small-to-medium enterprises. Its developer-friendly interfaces and on-demand pricing reduces barriers to entry for smaller clients, contrasting with hyperscalers’ enterprise-focused models.

Lambda Labs leverages Nvidia GPUs and high-performance networking to support training and inference for mid-scale models. Its focus on accessibility has made it a preferred choice for academic institutions and AI startups, enabling rapid prototyping and experimentation without the burden of long-term commitments.

The company’s smaller scale limits its ability to secure large GPU allocations or compete for enterprise contracts, but its agility positions it well for niche segments, particularly in democratising access to AI infrastructure for emerging players.

Nebius

Spun off from Yandex, Nebius has rapidly established itself as a distinctive player in the AI infrastructure market, operating a full-stack AI cloud with 30,000 Nvidia H200 GPUs as of March 2025 and plans to deploy over 22,000 Blackwell GPUs. Its platform, rebuilt in October 2024, integrates in-house hardware, energy-efficient data centres, and software tools like AI Studio – an inference-as-a-service offering with token-based pricing that has attracted nearly 60,000 users since launch.

Energy efficiency is one key differentiator for Nebius. Its Finland data centre features server-heat recovery systems and one of the world’s lowest power usage effectiveness (PUE) ratings, significantly reducing energy costs. The company’s proprietary servers, designed to operate at higher temperatures (40°C) with air cooling, lower maintenance costs and enable faster deployment.

On the software side, Nebius offers Soperator, a SLURM-based workload manager that optimises job scheduling, while its MLOps suite streamlines the AI lifecycle. Targeting developers, AI labs, and enterprises, Nebius is expanding in Europe and the U.S., with a focus on data sovereignty, which could be a strategic advantage in regulated markets.

Cerebras

Cerebras Systems, which is in the process of going public, takes a fundamentally different approach with its Wafer-Scale Engine (WSE), a massive chip integrating compute, memory, and interconnects for AI training and inference. Unlike GPU-based providers, Cerebras offers up to 10x faster training for certain LLMs compared to GPU clusters by eliminating much of the communication overhead inherent in distributed systems.

Its CS-3 system, powered by the third-generation WSE, competes with Nvidia’s DGX clusters for frontier AI research, while its Condor Galaxy supercomputers provide exascale compute for global AI development. Cerebras’ integrated hardware and software approach simplifies model development but targets a premium segment focussed on ultra-large models.

The current market heavily favours providers with access to Nvidia’s high-end GPUs, which are in short supply due to their dominance in AI workloads. This scarcity gives CoreWeave, Nebius, and Lambda Labs a competitive edge, as their GPU allocations attract clients unable to secure chips directly. However, the rise of alternative AI accelerators – like Cerebras’ WSE, AMD’s Instinct MI300, Google’s TPUs, and AWS’s Trainium – could disrupt this dynamic. If these accelerators match or surpass Nvidia’s performance, demand for Nvidia GPUs could slow, favouring hyperscalers with broader infrastructure, diverse chip portfolios, and greater financial resources to integrate alternatives. Hyperscalers’ economies of scale and established client bases could outpace specialised providers in a less supply-constrained market. Additionally, the capital intensity of scaling GPU clusters and data centres remains a significant barrier, requiring substantial debt financing and exposing providers to financial risks –  especially if market dynamics shift.

Further, it is important to consider the broader risks facing these specialised AI infrastructure pure plays. Demand uncertainty looms, as long-term monetisation of AI applications remains unproven, potentially leading to industry consolidation around hyperscalers if GPU supply normalises. Customer concentration, such as CoreWeave’s reliance on two clients for 77% of 2024 revenue, introduces volatility. Capital intensity for expanding infrastructure, coupled with geopolitical factors like U.S.-China chip export restrictions and data sovereignty laws, could disrupt supply chains. 

Energy constraints, too, are becoming critical, with AI’s massive power demands straining grids, necessitating sustainable solutions. These hurdles require specialised providers to innovate continuously while navigating a competitive, capital-intensive environment.

Sizing the opportunity 

According to industry analysts, the AI infrastructure market is likely to experience rapid growth, driven by surging demand for high-performance computing and widespread AI adoption. Efficiency gains are accelerating this expansion, as lower costs enable broader deployment. This phenomenon is known as the Jevons Paradox – the idea  that increased efficiency drives higher resource consumption. 

As shown in Figure B, Precedence Research projects that “the global artificial intelligence infrastructure market size will grow from US$47 billion in 2024 to USD 499 billion by 2034 (CAGR of 26.60%). North America currently leads with US$19.36 billion in 2024, driven by innovation, mergers, and hyperscaler investments. The report highlights particularly strong growth in inference infrastructure, projected at a 31% CAGR compared to 23% for training infrastructure, reflecting the broader deployment of AI applications across sectors.

Figure B: The AI Infrastructure Market Opportunity

Source: Precedence Research, February 2025.

Morgan Stanley, in a recent report, offers an even more aggressive outlook, projecting cumulative AI infrastructure spending exceeding US$3 trillion by 2028, with US$2.6 trillion allocated to data centers (chips and servers).

Further, it appears we are approaching what could be called an “Inference Inflection Point” in the market. While media attention and investment have concentrated on training infrastructure –  driven by GPU scarcity and the prestige of foundation model development – the long-term volume opportunity is increasingly shifting toward inference. 

This inflection point could significantly shift competitive dynamics in the coming years. Providers with global distribution, energy-efficient designs for sustainable scaling, and flexible pricing models may capture disproportionate value as inference workloads explode. For investors, this suggests looking beyond today’s training-dominated headlines to identify companies well-positioned for the expanding inference market that will follow the current training-focussed boom.

We can already see strategic positioning for this shift among the major players. CoreWeave, while currently focused on training contracts, is expanding its inference capabilities. Nebius has aggressively developed its AI Studio platform specifically targeting inference workloads with token-based pricing. Even Cerebras, despite its high-end training focus, is exploring how its architecture can serve certain inference applications efficiently.

Conclusion

AI infrastructure is a cornerstone of the intelligence era, enabling transformative technologies through specialised hardware, advanced connectivity, and tailored software. Its evolution from traditional cloud computing marks a pivotal shift, driven by the demands of training and inference. 

Specialised providers such as CoreWeave, Lambda Labs, Nebius, and Cerebras are redefining the landscape with innovative designs and full-stack solutions, but success hinges on several critical factors.

Key considerations for the sector’s future include:

• The pace of AI adoption and the scalability of infrastructure solutions across industries.
• Innovations in energy efficiency, connectivity, and hardware to address power and performance constraints.
• Competition between hyperscalers and specialised providers, differentiated by pricing, software, and niche expertise.
• Geopolitical and regulatory developments impacting supply chains and market access.

The likely winners will be companies that balance technological innovation with practical scalability, addressing enterprise needs while navigating capital and energy constraints. The opportunity extends beyond infrastructure providers to the broader ecosystem of semiconductor, software, and edge computing players. 

AI infrastructure represents a transformative force, but its path forward requires careful analysis of technological capabilities, market dynamics, and disciplined execution.

At AlphaTarget, we invest our capital in some of the most promising disruptive businesses at the forefront of secular trends; and utilise stage analysis and other technical tools to continuously monitor our holdings and manage our investment portfolio. AlphaTarget produces cutting edge research and our subscribers gain exclusive access to information such as the holdings in our investment portfolio, our in-depth fundamental and technical analysis of each company, our portfolio management moves and details of our proprietary systematic trend following hedging strategy to reduce portfolio drawdowns.  To learn more about our research service, please visit https://alphatarget.com/subscriptions/.

Chinese company DeepSeek has taken the AI world by storm with their recent unveiling of cutting-edge large language models (LLMs), specifically DeepSeek-V3 and its reasoning-focused variant DeepSeek-R1. These open-source models demonstrate performance comparable to leading competitors (Figure A) at roughly 1/10th the training cost and significantly lower inference cost. In simple terms, training cost refers to the expenses incurred in developing and fine-tuning an AI model, including the computational resources, data preparation, and model optimisation required to build the model. On the other hand, inference cost refers to the ongoing costs of running the trained model to make predictions or decisions i.e. when ChatGPT answers users’ queries. 

By releasing its models under an open-source license, DeepSeek allows other organisations to replicate and build upon its work. This democratises access to advanced AI technologies, enabling a broader range of entities to develop and deploy AI solutions at a much lower cost. The development raises fundamental questions about the economics of AI: While improved efficiency could accelerate AI adoption, it challenges assumptions about near-term semiconductor demand for training capex, cloud infrastructure spending, and the distribution of value across the technology landscape. 

In this article, we examine DeepSeek’s key technical innovations, explain how they achieved frontier model performance despite US chip export restrictions, and analyse the broader implications for the technology industry. 

Figure A: Benchmark Performance of DeepSeek-R1

Source: DeepSeek-R1 Technical Report

Market reaction and initial impact 

The US stock market awakened to DeepSeek’s disruption in the third week of January. That coincided with the release of DeepSeek-R1 (on 20 January 2025). The model, released under the MIT licence along with its source code, demonstrated exceptional capabilities in reasoning, mathematics, and coding tasks, matching the performance of OpenAI’s o1 “reasoning” model at a fraction of the cost. DeepSeek complemented this release with a web interface for free access and launched an iOS application that quickly reached the top of the App Store charts.

Given the US export ban on top-end GPUs to China, which many argued was necessary for the US to maintain its lead in the AI race, the market was shocked to see a China-based frontier model rival the capability of US-based models! By 27 January, as news of DeepSeek’s breakthrough gained widespread attention, technology stocks experienced sharp declines, with Nvidia falling 16%, Oracle 12% and smaller AI infrastructure stocks declining 20-30% in a single trading session!

While R1’s release finally got the market’s attention, the foundation was laid a month earlier when DeepSeek-V3 was released around Christmas. On 26 December 2024, Andrej Karpathy (formerly Director of AI at Tesla and founding member of OpenAI) highlighted the remarkable achievement: DeepSeek-V3 had reached frontier-grade capabilities using just 2,048 GPUs over two months – a task that traditionally required clusters of 16,000+ GPUs (Figure B). This efficiency gain fundamentally challenged assumptions regarding the computational resources required for developing advanced AI models.

Figure B: Andrej Karpathy on DeepSeek-V3

Source: X 

Technical innovations

DeepSeek’s efficiency gains arise from several key architectural innovations that fundamentally rethink how large language models process and generate text.

At its core, DeepSeek employs a Mixture-of-Experts (MoE) approach – similar to relying on the most relevant specialist instead of the entire team of doctors when treating a patient. By selectively activating only 37B of 671B parameters for each piece of text (token), the system achieves remarkable efficiency gains. For perspective, Meta’s Llama 3 405B required 30.8M GPU-hours, while DeepSeek-V3 used just 2.8M GPU-hours – an 11x efficiency gain. At an assumed cost of US$4 per GPU hour, this translates to roughly US$11.2 million in training costs versus US$123.2 million for Llama 3 405B.

There are several technical innovations worth calling out that focus on efficiency, including:

• Multi-head Latent Attention (MLA): MLA compresses the information required in the “attention” mechanism through low-rank projections. MLA is analogous to a very good summary of a large document. This technique drastically reduces memory demands during inference while retaining the full performance of standard multi-head attention.
• Multi-Token Prediction (MTP): Rather than generating one word (token) at a time, MTP attempts to predict multiple future words simultaneously. This achieves an 85-90% success rate in predicting upcoming tokens, resulting in 1.8x faster text generation.
• Auxiliary-loss-free load balancing: Traditional MoE architectures struggle to balance workload across the “experts.” This is akin to a medical emergency department where some doctors are overworked versus others! DeepSeek-V3 introduced a novel “dynamic bias adjustment” mechanism that ensures balanced expert workloads without compromising performance, achieving 90% expert utilisation.

Prior to DeepSeek’s arrival, many of these architectural and design choices were already known in the research community. For instance, it was well understood that the MoE approach yields 3x to 7x efficiency gains compared to dense models. Nonetheless, DeepSeek managed to push the boundaries of efficiency even further without compromising model quality. By building DeepSeek efficiently without relying on state-of-the-art GPUs, the company demonstrated that necessity is the mother of invention

Reinforcement learning breakthrough

DeepSeek’s R1 technical paper reveals a significant step change in how reasoning capabilities can be developed in LLMs. The widely accepted standard had been to use supervised learning with human-curated datasets followed by reinforcement learning with human feedback (RLHF). DeepSeek, however, showed that sophisticated reasoning can emerge primarily through reinforcement learning alone. In its technical report, DeepSeek notes that its work is “the first open research to validate that reasoning capabilities of LLMs can be incentivized purely through RL, without the need for SFT [Supervised Fine-Tuning].”

The significance of this approach lies in its ability to transcend the limitations of human-demonstrated problem-solving patterns. Traditional supervised learning methods can only replicate reasoning strategies present in the training data. In contrast, DeepSeek’s reinforcement learning approach allows the model to discover novel problem-solving strategies through systematic exploration. This is analogous to the difference between learning chess by studying grandmaster games versus learning by playing millions of games and discovering new strategies independently.

Overcoming the US chip ban

DeepSeek’s technical reports also shed light on how the company navigated the U.S. export ban on top-end GPUs like the Nvidia H100. The H800 GPUs available (legally) to DeepSeek significantly reduced NVLink link bandwidth and double-precision computing capabilities. To overcome these limitations, DeepSeek implemented optimisations such as:

• Restricting token processing to groups of 4 GPUs, thus minimising data transfer bottlenecks.
• Developed techniques to handle internal (NVLink) and external (InfiniBand) communication concurrently.
• Implemented FP8 mixed-precision training, halving memory requirements compared to traditional approaches.
• Developed custom kernel software for efficient local expert forwarding.

Investment implications

DeepSeek’s innovations will potentially upend how value is created and captured in the AI industry. In this section, we consider some of the potential ramifications of DeepSeek’s innovations, cost efficiency, and release under an MIT license.

At a high-level, DeepSeek’s innovations and cost efficiency are likely to force a re-think on the ravenous appetite for high-end AI accelerators. However, technology history suggests that improved efficiency often leads to increased total resource consumption — a phenomenon known as the Jevons paradox. As barriers to entry fall, more organisations can experiment with and deploy AI solutions, potentially driving higher aggregate demand for AI infrastructure. 

Naturally, for semiconductor companies, especially those focused on AI acceleration, these trends present both challenges and opportunities. DeepSeek’s ability to produce a top-grade frontier LLM without access to NVIDIA’s high-end GPUs suggests that some companies might have been complacent with respect to engineering and architectural optimisations, and have instead chosen to use a sledgehammer capex approach to develop increasingly sophisticated models. It is possible that DeepSeek’s success will cause some rethink on this front, which might in turn moderate high-end GPU demand for training capex. This means that companies might be able to do more with their existing GPUs and their useful lives might even be extended. However, the fundamental dynamics of AI development remain compelling: As long as scaling laws hold – where model performance improves as a power-law function of size, dataset, and compute resources – we can expect compute-hungry AI algorithms to require ever more processing power to solve increasingly challenging problems. Moreover, as AI models become more widely available, the demand for inference-optimised semiconductors (inference capex) is likely to grow substantially.

Major cloud providers’ current capital expenditure plans suggest confidence in this longer-term vision despite efficiency gains. Microsoft has reiterated its US$80 billion capex commitment, while Alphabet projects approximately US$75 billion in CapEx for 2025, up significantly from US$52.5 billion in 2024. Amazon expects to maintain its Q4 2024 quarterly investment rate of US$26.3 billion through 2025. While it is potentially too early to see strategic shifts in response to DeepSeek’s innovations, these investment levels suggest hyperscalers anticipate growing demand for AI compute, and particularly inference workloads. Notably, hyperscalers are positioning themselves as model-agnostic platform providers – evidenced by Microsoft making DeepSeek’s R1 available on GitHub and Azure AI Foundry despite its close partnership with OpenAI, and Amazon integrating DeepSeek R1 into its Bedrock and SageMaker platforms.

This evolution signals a broader shift in competitive advantage from raw model capabilities towards proprietary data assets, distribution channels, and specialised applications. Enterprise software businesses, which typically focus on building applications around models rather than developing models themselves, stand to benefit from the widespread availability of high-quality open-source foundation models. As these models improve and become more accessible, competitive advantage will increasingly derive from unique data assets and distribution channels. Enterprises with specialised data repositories, such as Salesforce (CRM data) or Bloomberg (financial data), could develop highly targeted AI agents that leverage their proprietary data advantages.

The platform landscape presents varying implications for different players. Meta appears well-positioned to benefit from LLM commoditisation (a strategy it has followed with its own Llama models), as it can deploy AI innovations to enhance content discovery and user engagement across its vast social networks. Alphabet presents a more complex case: while its Google search advertising model faces potential disruption from AI-powered “answer engines,” its YouTube platform and cloud business could benefit substantially from increased AI adoption and deployment. 

Obviously, there is a question with respect to where one would undertake the inference task. DeepSeek’s approach to providing “distilled” models (1.5B–70B parameters) opens new possibilities for edge computing. These smaller models enable resource-constrained devices to leverage advanced AI capabilities while navigating practical constraints like battery life and thermal limitations. Apple’s current hybrid approach — running on-device LLMs for certain tasks while routing others to private cloud infrastructure built using Apple’s M-series silicon – could become the standard blueprint, particularly in consumer applications. This hybrid model effectively balances privacy considerations and response times with computational capabilities.

As the industry evolves, significant uncertainties remain regarding the distribution of economic value between frontier model developers, infrastructure providers, platform companies, and application developers. While investors will need to keep a close eye on these shifting value propositions, and the winners and losers are likely to emerge only in the fullness of time, one aspect appears increasingly clear: We are entering a fascinating phase of AI development. The improved efficiency and accessibility demonstrated by innovations like DeepSeek could democratise access to AI capabilities, potentially making businesses and consumers the ultimate beneficiaries of this technological revolution.

Conclusion

DeepSeek’s innovations are not just a technical achievement; they signal a potential restructuring of the AI industry’s competitive dynamics. By achieving frontier model performance at roughly 1/10th the traditional training cost and releasing the model under the MIT license, DeepSeek has challenged fundamental assumptions about AI development and deployment. 

In this article, we have explored three key implications of this development. First, we have considered the balance between training and inference capex through the lens of Jevons paradox which suggests that lower barriers to entry could drive higher aggregate compute consumption over time. Second, we have considered the potential shift in value creation from raw model capabilities towards data assets, distribution platforms, and specialised applications. Finally, we have contemplated how the emergence of efficient “distilled” models enables new deployment paradigms, from edge computing to hybrid architectures, potentially expanding AI’s practical applications.

For investors, there are many nuances to consider including:

• The near-term risks versus medium to long-term opportunities for semiconductor businesses.
• The position of hyperscalers as model-agnostic compute and inference platforms. 
• Democratisation of AI models and its effects on enterprise software businesses.
• The pros and cons for platform companies with access to unique data and/or strong distribution channels.

While the geopolitical implications of DeepSeek are still unfolding, the development clearly demonstrates that multiple paths to AI progress exist beyond simply scaling up computing resources and that a path to more democratised AI models is likely feasible. We believe organisations across the spectrum will be reassessing their AI strategies in light of these developments, with perhaps their lens focused on “value addition” and “Return on Investment” rather than raw model size or computing power. In a nutshell, DeepSeek is excellent news for the users of AI and by lowering both the training and inference costs, it has massively increased the Return on Investment (ROI) on AI.

At AlphaTarget, we invest our capital in some of the most promising disruptive businesses at the forefront of secular trends; and utilise stage analysis and other technical tools to continuously monitor our holdings and manage our investment portfolio. AlphaTarget produces cutting edge research and our subscribers gain exclusive access to information such as the holdings in our investment portfolio, our in-depth fundamental and technical analysis of each company, our portfolio management moves and details of our proprietary systematic trend following hedging strategy to reduce portfolio drawdowns. 

To learn more about our research service, please visit: https://alphatarget.com/subscriptions/.

Agentic artificial intelligence (AI) systems are fast-emerging as the next leap forward in AI systems development. Unlike traditional generative AI, which excels at creating content and responding to specific queries, agentic AI takes this capability further by enabling autonomous decision-making and proactive task execution within a given environment. This shift from passive content creation to active, decision-making AI has the potential to revolutionise how businesses operate. At a high level, while generative AI creates, agentic AI acts.

What is an AI agent? 

An AI agent is a software program (Figure A) designed to interact with its environment, gather data, and perform tasks autonomously to achieve predefined goals. While humans define these goals, the AI agent independently determines the best course of action to fulfill them. 

For instance, in a contact centre, an AI agent can resolve customer queries by asking questions, retrieving information from internal resources, and providing solutions. Based on the customer’s responses, it decides whether to resolve the query itself or escalate it to a human representative.

AI agents function as rational agents that can analyse data and their environment to make informed, goal-oriented decisions. The following key capabilities set them apart from traditional software:

• Perception and Data Gathering: AI agents use interfaces to sense their environment. For instance, a robotic agent collects data through physical sensors, while a chatbot interprets text input from customer queries.
• Decision-Making: Using the collected data, AI agents predict the best outcomes and determine the next steps to achieve their objectives. For example, a self-driving car navigates obstacles using sensor inputs and pre-trained models.
• Collaboration: AI agents can collaborate with other agents and humans, in order to complete multi-step complex tasks.
  
  

Figure A: AI agent architecture

Source: Falkor DB

Benefits of AI

AI agents bring significant advantages to businesses and customers alike by streamlining operations and enhancing experiences. Key benefits include:

1. Enhanced Productivity

AI agents streamline operations by automating tasks, reducing inefficiencies, minimising human errors, and eliminating the need for manual effort, resulting in lower operational expenses. Their ability to adapt to evolving conditions ensures steady performance, further enhancing resource efficiency.

2. Cost Efficiency

AI agents automate workflows, cutting down inefficiencies, reducing human errors, and eliminating manual tasks, which lowers operational costs. Their flexibility allows them to maintain consistent performance, even in dynamic conditions, further enhancing resource utilisation.

3. Improved Decision-Making

Sophisticated AI agents utilise machine learning (ML) to process vast amounts of real-time data. This capability enables businesses to make accurate predictions and data-driven decisions swiftly. For example, AI agents can analyse market trends to optimise product positioning during an advertising campaign.

4. Superior Customer Experience

AI agents enhance customer engagement by delivering personalised interactions, prompt responses, and tailored recommendations. Businesses can boost customer satisfaction, conversion rates, and loyalty through these advanced, data-driven insights.

Given the unique capabilities of AI agents, the world’s transition to using active AI agents could prove to be an inflection point in how software is built and consumed. Currently, several major software businesses are evolving their technology stacks with an AI-first strategy. 

For instance, enterprise giant Microsoft has integrated AI agents across its enterprise suite through its Copilot initiative, with capabilities ranging from automated coding, email summarisation, intelligent scheduling and improved search capabilities. 

Similarly, IT workflow automation company ServiceNow is deploying specialised AI agents for IT service management, customer service, HR, and procurement workflows. These agents can handle everything from IT ticket resolution to employee onboarding, learning from each interaction to improve their performance while maintaining human oversight for critical decisions.

AI agents: A brief history

The concept of AI agents is not entirely new – it has roots in the “expert systems” of the 1960s and 1970s. These early attempts at AI used rule-based programming to mimic human decision-making in specific domains.

MYCIN, developed at Stanford in the 1970s, was one of the first expert systems. Its inference engine could diagnose blood infections and recommend antibiotics, using approximately 600 predefined rules. While groundbreaking for its time, these rule-based systems were inherently limited by their inability to learn from experience or handle novel situations.

Arguably the biggest step change enabling modern AI agents came with two more recent significant advances in artificial intelligence. First, deep learning breakthroughs in the 2010s enabled systems to learn from vast amounts of data and improve their performance over time. Then, the development of large language models (LLMs) in the early 2020s provided the sophisticated reasoning and natural language capabilities that make today’s AI agents possible.

These foundation models dramatically improved natural language processing capabilities, enabling AI agents to understand nuanced instructions, generate contextually appropriate responses, and engage in more natural interactions. This advancement means agents can now interpret complex business requests and communicate their actions clearly to users.

Another crucial innovation is retrieval-augmented generation (RAG), which allows AI agents to combine their built-in knowledge with real-time access to enterprise data and documents. This technology enables agents to ground their responses in an organisation’s specific context, policies, and procedures while maintaining accuracy and relevance. RAG is a key tool in preventing LLMs from “hallucinating” – that is, generating misleading or inaccurate results.

Overall, this progression from simple rule-based systems to today’s sophisticated AI agents represents more than incremental improvement – it marks a fundamental shift in how organisations can deploy AI. Rather than following predetermined rules, modern agents can understand context, learn from experience, and autonomously pursue objectives while operating within appropriate business constraints.

AI agents in action

Major technology companies are making significant investments in agentic AI as the convergence of advanced language models, deep learning capabilities, and enterprise integration frameworks has created new opportunities for business transformation. In terms of the business case, there are opportunities with respect to cost efficiency, enhanced productivity, and improvements in outcomes, as demonstrated by a number of early deployments:-

Automating internal operations: At telecommunications giant Deutsche Telekom, an AI agent named “askT” is transforming the company’s internal operations. Serving approximately 10,000 employees weekly, askT not only answers questions about internal policies and benefits but is also actively managing administrative tasks like leave applications. In addition to cost efficiencies, the system provides consistent, 24/7 service availability while continuously learning from interactions to improve its responses.

Reimaging financial analysis: In the financial sector, organisations are leveraging AI agents’ sophisticated analytical capabilities. Moody’s, for instance, has implemented AI agents to analyze vast amounts of market data and company financials, enabling more comprehensive risk assessments. Moody’s system consists of a network of 35 specialized AI agents, each designed for specific analytical tasks and working within a coordinated multi-agent system. According to Nick Reed, Chief Product Officer at Moody’s, these agents have transformed how the company conducts crucial research tasks that were previously outsourced to lower-cost regions, such as industry benchmarking and SEC filing reviews.

Customer experience innovation: eBay’s implementation of AI agents is another illustration of the technology’s potential impact across multiple business functions. The company’s sophisticated “agent framework” orchestrates multiple LLMs for various tasks such as assisting with customer service inquiries, writing code,  and creating marketing campaigns. The system learns from employee interactions to understand specific preferences and work styles, enabling increasingly autonomous operation. eBay is considering expanding this capability to help buyers find items and assist sellers with listings, showcasing how AI agents can enhance core business operations while improving the customer experience.

The SaaS disruption debate 

Microsoft CEO Satya Nadella made a rather provocative claim during his appearance on the BG^2 podcast with hedge fund manager and venture capitalists Brad Gerstner and Bill Gurley. Nadella noted that “the business logic is all going to these AI agents” and suggested that they would fundamentally transform how organisations interact with their enterprise Software as a Service (SaaS) applications. This viewpoint challenges the traditional SaaS model that has dominated enterprise software for the past two decades. 

Currently, business applications like CRM, ERP, HCM, and IT workflow systems primarily function as sophisticated data management tools, operating on a CRUD (Create, Read, Update, Delete) model with humans directing most operations. Nadella’s vision suggests a fundamental shift where the intelligence layer of these systems will migrate from isolated applications to a unified AI layer in the technology stack.

In a utopian world with no accumulated technology debt, one could imagine a single AI layer seamlessly handling almost all business processes and functionalities. This scenario is not entirely theoretical; tech-savvy startups might build their systems around an AI-first architecture, potentially bypassing traditional business applications in favor of a more streamlined, AI-driven approach with simple database backends.

However, the reality is more complex. Enterprises have valuable data locked within various legacy systems and platforms, therefore realising a unified AI tier that can seamlessly orchestrate all business operations is extremely challenging. Furthermore, the incumbent leaders are not standing still. For instance, ServiceNow’s AI agents are being designed to work across previously siloed workflows, from IT service management to HR processes. Similarly, Microsoft is integrating Copilot capabilities throughout its enterprise suite, while Salesforce’s Einstein GPT and AI Cloud are reimagining customer relationship management and employee interactions through the lens of AI-driven automation.

Given the current state-of-play, predictions about the death of SaaS appear exaggerated and we expect many SaaS players to evolve by incorporating AI features. Companies that can effectively bridge the gap between legacy systems and AI-driven operations stand to capture significant value in the enterprise market. Infrastructure providers supporting AI operations, particularly those offering solutions for data integration and AI orchestration, could see increased demand. Additionally, new entrants might emerge with AI-first approaches that challenge traditional enterprise software categories entirely. The winners in this transition will likely be those who can effectively combine deep domain expertise with advanced AI capabilities while addressing the practical challenges of enterprise data integration.

Sizing the opportunity 

The emergence of AI agents represents part of a broader technological transformation that analysts believe could rival the mobile and cloud computing revolutions in scope and impact. According to Battery Ventures’ 2024 State of the OpenCloud report, the AI software TAM is worth ~US$4 trillion as it disrupts traditional software, services, and labour markets (Figure B).

Figure B: The Massive AI Market Opportunity

Source: Battery Ventures, State of the OpenCloud Nov 2024

The precise market sizing for agentic AI remains unclear due to the emergent nature of this technology. However, we note that the impact of AI agents extends beyond direct market size. Cloud providers are making unprecedented investments in AI infrastructure to support these emerging technologies. This infrastructure buildout provides the foundation for widespread adoption of AI agents across industries.

Going forward, it will be important to watch the pace of agentic AI technology adoption within enterprises as well as regulatory developments. While some regulations may stifle deployment, ultimately regulation developed in conjunction with industry leaders and other key stakeholders will likely create a more stable environment for long-term investment.

Conclusion

Agentic AI is a fundamental shift in enterprise technology moving beyond simple automation to create systems capable of autonomous, complex decision-making. While early adoption shows promising results, success will depend on several critical factors.

For investors evaluating opportunities in this space, key considerations include:

• The pace of enterprise adoption and deployment success rates
• Regulatory developments, particularly around autonomous decision-making
• Competition between incumbent SaaS providers and AI-first startups
• Infrastructure requirements and associated investment opportunities

The winners in this transition will likely be companies that can effectively bridge the gap between legacy systems and AI-driven operations, while addressing the practical challenges of enterprise data integration. We believe both established players as well as new AI-first players can win in this market. The opportunity also extends beyond direct AI agent providers to the broader ecosystem of companies providing essential infrastructure, security, and integration services. For investors, agentic AI represents a potentially transformative opportunity – but one that requires careful analysis of technological capabilities, business models, and go-to-market strategies.

At AlphaTarget, we invest our capital in some of the most promising disruptive businesses at the forefront of secular trends; and utilise stage analysis and other technical tools to continuously monitor our holdings and manage our investment portfolio. AlphaTarget produces cutting edge research and our subscribers gain exclusive access to information such as the holdings in our investment portfolio, our in-depth fundamental and technical analysis of each company, our portfolio management moves and details of our proprietary systematic trend following hedging strategy to reduce portfolio drawdowns.  To learn more about our research service, please visit https://alphatarget.com/subscriptions/.