The convergence of the end of Moore’s law[1] with the surge of generative AI has opened up significant investment opportunity across the semiconductor ecosystem. When shrinking the transistor can no longer carry the load, new innovation in architecture, memory, interconnect, packaging and power have the potential to redefine an entire industry.
Executive Summary
Every major technology shift of the last half century, from the personal computer to the smartphone to the cloud, has been built on semiconductors. Artificial intelligence is no different, except in scale. Every model trained, every question answered by a chatbot, and every task completed by an AI agent runs on chips, and demand for those chips is growing faster than at any point in the industry’s history. The global semiconductor market is expected to grow from $1.7 trillion in 2026 to $3.2 trillion by 2030[2]
The world’s most valuable companies either make semiconductors or design their own. The largest hyperscalers are expected to spend more than $800 billion on capex this year, growing to $3 trillion by 2030[3], most of it being spent on chips and the systems around them. The recent fall in AI token prices strengthens that case: cheaper tokens drive more usage, and every token still takes silicon, memory, and power to produce. Semiconductors are no longer a component that technology companies buy. They are the foundation that the next era of computing will be built on.
The investment opportunity in this sector extends well beyond the handful of public companies that dominate today’s headlines. AI has exposed new bottlenecks across the entire system: how data moves between chips, how memory keeps up with compute, and how power reaches the processor. Each is a large market in its own right (with TAMs ranging from $26 billion to $837 billion+). The current incumbents are not necessarily the companies best placed to solve these issues. They have roadmaps to deliver and publicly monitored bottom lines that tend to limit their R&D risk tolerance. New solutions are increasingly coming from startups, which is fitting: Silicon Valley, after all, was named for the semiconductor startups that drove its evolution.
I. A Decade of Consolidation and Underinvestment (2010-2022)
Semiconductor venture capital became an overlooked asset class
Despite the semiconductor industry’s central role in the global economy, investment in early-stage semiconductor companies remained disproportionately small relative to the sector’s importance. The core problem stemmed from a misconception dating back decades, to the start of Silicon Valley and the original semiconductor business model.
In 1957, the “Traitorous Eight” engineers left Shockley Semiconductor Laboratory to start Fairchild Semiconductor, backed by industrialist Sherman Fairchild and financing arranged by investment banker Arthur Rock. Fairchild pioneered the Silicon Valley model, demonstrating that a small group of engineers could leave an established company, raise outside capital, and build a new technology business. Rock would later partner with Tommy Davis to establish Davis & Rock in 1961, one of the earliest dedicated venture capital firms on the West Coast, helping formalize a model in which investors provided capital to ambitious technology entrepreneurs in exchange for equity[4]
The impact of Fairchild extended far beyond the company itself. Its engineers and executives went on to found or build many of the next generation of semiconductor companies, including Intel and AMD, while Fairchild alumni also became some of the earliest architects of Silicon Valley’s venture capital ecosystem. Eugene Kleiner, one of the original Fairchild founders, went on to co-found Kleiner Perkins with Tom Perkins, while Don Valentine, another Fairchild employee, founded Sequoia Capital. This created a powerful cycle in which semiconductor entrepreneurs became venture capitalists, venture capitalists funded new semiconductor companies, and those companies produced the next generation of entrepreneurs and investors. For its first few decades, venture capital and semiconductors were essentially synonymous.
There was, however, one problem building underneath the surface: the early semiconductor model came with a heavy price tag. Early chip companies were integrated device manufacturers (IDMs), designing, manufacturing, and often packaging their own chips in their own facilities. Owning a fab was treated as a requirement for competing (AMD founder Jerry Sanders famously declared that “real men have fabs.”). For as long as fabs were relatively affordable, the model worked. But each new generation of smaller transistors required more advanced equipment, cleaner facilities, and more complex processes. The cost of building a fab rose so steadily that it earned its own rule of thumb, Rock’s Law, which holds that the cost of a new fab doubles roughly every four years. Today a single leading-edge fab costs more than $20 billion.
As costs climbed, manufacturing at the leading edge consolidated into the hands of a few companies that could afford it. Even large chipmakers struggled to keep up: AMD itself spun off its manufacturing operations in 2009, forming what became GlobalFoundries. For startups, the IDM model was effectively out of reach. Investors came to view semiconductors as an industry where every new company needed hundreds of millions or billions of dollars in factories before it could sell a product, and that view hardened into conventional wisdom. Although misconception lingers, the IDM model is no longer how most of the industry functions today. When Taiwan Semiconductor Manufacturing Company (TSMC)[5] pioneered the pure-play foundry model in 1987, it separated chip design from chip manufacturing. A foundry spreads the enormous cost of a fab across hundreds of customers, giving even a small design team access to world class manufacturing. The rest of the industry has since specialized around this model. Design tools come from EDA vendors like Synopsys and Cadence, core IP can be licensed from companies like Arm, manufacturing is outsourced to foundries like TSMC and GlobalFoundries, and packaging and testing are handled by specialized OSAT (Outsourced Semiconductor Assembly and Test) partners like ASE or JCET. Many of the most valuable semiconductor companies in the world, including NVIDIA, Qualcomm, and Broadcom, built their businesses without ever owning a leading-edge fab.
For a venture-stage chip company, this changes the economics entirely. Rather than spending on factories and equipment, a startup spends its capital primarily on engineering talent, design tools, and tape outs. It can validate its technology through shared “multi-project wafer” runs, which split the cost of a fab run across several customers, before committing to a full production run. Many startups also target mature process nodes, where manufacturing costs are a fraction of the leading edge. In practice, today’s semiconductor startups are far more capital-efficient than the market tends to believe. For generalist VCs, this misconception, along with the deep technical diligence required to assess semiconductor opportunities, drove investors to other areas within tech such as software. The resulting capital gap was substantial. From 2013 through 2022, semiconductor startups received less than 1% of U.S. venture investment per year[6]
The semiconductor industry became increasingly concentrated
Between 2010 and 2020, the number of publicly traded semiconductor companies in the United States declined from 160 to 97, reflecting a prolonged period of industry consolidation[7] . Large companies expanded their scale and created value through acquisitions, and as a result, a relatively small number of scaled semiconductor companies controlled an increasing share of the industry’s resources, customer relationships, manufacturing access, talent, influence, and intellectual property.
As consolidation reduced the number of independent mid-sized semiconductor companies, the pool of readily available acquisition targets became more limited. At the same time, large public-to-public transactions became increasingly difficult to execute because of regulatory and geopolitical considerations. NVIDIA’s proposed $40 billion acquisition of Arm, for example, was terminated in February 2022 following significant regulatory opposition. Earlier attempts by Broadcom to acquire Qualcomm for $117 billion in 2017 and Qualcomm to acquire NXP for $47 billion in 2016 similarly faced significant regulatory scrutiny. For the largest chipmakers, buying another scaled public company was no longer a reliable path to growth. The M&A landscape shifted: instead of growing revenue and market share by acquiring peers, incumbents shifted focus toward acquiring capabilities they could not efficiently build themselves in order to grow revenue and market share.
Increasingly constrained engineering talent pipeline
The industry’s consolidation and underinvestment also contributed to a longer-term challenge: the availability of experienced semiconductor engineering talent. Chip development depends on deep expertise in digital and analog design, architecture, verification, process engineering, packaging, photonics, and power electronics – skills that take years to build. With few startups being formed, engineers had fewer chances to gain hands-on company building experience and many left for higher paying jobs in other corners of tech. As a result, studies predict that the U.S. semiconductor workforce would need to grow by 115,000 workers by 2030 to keep up with the growing demand for semiconductor engineers. At existing degree completion rates, approximately 67,000 positions, or 58% of projected new jobs, could go unfilled. For startups, that makes experienced technical leadership scarce which has led to many multi-billion dollar acquihires of private semiconductor companies over the last five years.[8]
The resulting opportunity
With a lack of venture capital funding, industry consolidation, regulatory scrutiny and talent scarcity it is not surprising that investors have shied away. Taken together, we believe these factors have created a significant opportunity. Startups with differentiated technology and experienced teams are scarce, which has made each one more valuable. At the same time, capital that would typically compete to fund these opportunities has been deployed elsewhere.
As a result, the landscape was forming: an industry with high barriers to entry, rising global strategic importance and a white space for specialized venture capital firms. All that was missing was a catalyst…
II. Why 2023 Was the Inflection Point: The Convergence of AI, Architectural Innovation, and Strategic Demand
The industry just needed one catalyst, yet 2023 brought several. Moore’s Law, the industry’s reliable engine of progress, was sputtering just as ChatGPT turned generative AI into the largest surge in demand for compute the industry has ever seen. When shrinking the transistor can no longer carry the load alone, the work fell to everything around it: architecture, memory, interconnect, packaging, and power. Any one of these shifts would have been interesting, but together, they started a new investment cycle across the semiconductor ecosystem.
Moore’s Law stopped doing the heavy lifting
For decades, Moore’s Law (the observation that transistor counts doubling roughly every two years could drive performance gains, energy efficiency, and reduce costs) gave the industry a predictable roadmap: each generation of chips was smaller, faster, and more efficient.
That roadmap has become far more expensive and less rewarding. Leading-edge process development costs have soared, while each new node delivers diminishing performance and power gains. Effectively, Moore’s Law has hit a wall. As a result, performance improvement is moving from the transistor to the system: architecture, packaging, memory, interconnect, and power delivery have become the levers to improve performance.
This creates an opportunity for a broader range of semiconductor startups to develop technologies that can become critical components of much larger computing systems, and an area where the large incumbents are generally too risk-averse to derail current product roadmaps to explore.
ChatGPT triggered a new compute cycle – Training, Inference, and Agentic AI
The launch of ChatGPT from OpenAI in November 2022 turned generative AI from a research topic into a mass-market product that consumers and enterprises could use. The pace of adoption has been astounding:
- ChatGPT reached 1 million users within 5 days and its weekly active users have continued to grow, with 400 million users in February 2025 to 900 million in February 2026[9]
Competing products grew alongside it: Claude from Anthropic, Gemini from Google, Copilot from Microsoft, Muse from Meta. - Today, roughly 2.2% of U.S. households have paid AI subscriptions and that number continues to grow[10]
- Enterprise adoption has quickly followed. Enterprise spending on AI grew from $1.7 billion in 2023 to $37 billion in 2025[11]
Behind this surge in usage sits an enormous and growing demand for computing power, and it begins with how these models are built. Training is the process through which an AI model learns to understand and generate information. During training, an LLM is exposed to massive datasets and repeatedly predicts the next token, compares each prediction with the correct answer, and adjusts billions of internal parameters to improve its performance. Repeating this across trillions of tokens requires an immense volume of mathematical operations, typically run on tens of thousands of GPUs or custom accelerators working in parallel for weeks or months.
That requirement keeps rising: models are getting larger, datasets are expanding, and developers are applying more compute to improve reasoning and capabilities. Because the chips in a training cluster must exchange results constantly, training drives demand not just for processors but for high bandwidth memory, advanced packaging, optical interconnects that link thousands of chips into a single system, and the power infrastructure for facilities drawing hundreds of megawatts. Training is also not a one-time cost. Labs continually retrain and fine tune models to stay competitive, and every model that ships creates a second, ongoing wave of inference demand as millions of users engage with it.
If training is how a model learns, inference is how it does its job. Every time a user asks ChatGPT a question, summarizes a document, or generates code, the trained model runs a forward pass through its billions of parameters to produce a response, one token at a time. A single inference request uses far less compute than a training run, but inference never stops. It scales with every user, every query, and every application built on top of a model. As adoption has grown to hundreds of millions of weekly users and AI features are embedded in enterprise software, inference is becoming a larger share of total AI compute and, unlike training, a recurring cost that grows directly with usage.
Inference also places different demands on hardware. Response speed matters to users, so latency, memory bandwidth, and cost per token become the key metrics, rather than raw training throughput. This has opened the market to specialized inference chips and architectures beyond the GPU, and it drives demand for high bandwidth memory, low power networking, and efficient power delivery in data centers, as well as for edge devices that run smaller models locally.
The next phase of demand is Agentic AI: systems that do not simply answer a single prompt but plan and carry out multi-step tasks on a user’s behalf, such as researching a topic, writing and testing code, or completing a workflow across several applications. To finish one task, an agent may call a model dozens or hundreds of times, reasoning through intermediate steps, using tools, checking its own work, and keeping a large and growing context in memory. As a result, a single agentic task can consume many times the tokens of a traditional chatbot exchange. Agents can also run continuously in the background, and many can run in parallel for a single user or business. This multiplies inference demand rather than merely adding to it. It also increases demand for memory capacity to hold long contexts, for CPUs to orchestrate tools and workflows, and for fast interconnects and reliable power to keep large inference fleets running around the clock. Agentic AI is well on its path towards reaching the same enterprise adoption that chat-based AI has and could become the largest driver of AI compute demand over the next several years.
Investment inflection point
The industry could no longer rely on smaller transistors to deliver the gains customers expected, just as training, inference, and agentic AI began demanding more compute than anyone had planned for. When the transistor stops doing the heavy lifting, the bottlenecks move to the rest of the system: how fast data moves between chips, how much memory sits beside them, how densely they can be packaged, and how efficiently they can be powered. These constraints now limit AI performance as much as the processors themselves do, and each represents a large and fast growing market, as covered in more detail in the following sections.
For investors, this marks a structural shift rather than a cyclical upturn. For decades, value in semiconductors accrued mainly to whoever led the next process node, a race only a handful of companies could afford to run. Today, value is increasingly created at the system level, where a startup with a better laser, interconnect, packaging approach, or power converter can become a critical component of a multibillion-dollar AI deployment. Incumbents are often slow to pursue these technologies because doing so would disrupt existing product roadmaps, which leaves the door wide open for startups. In practice, large incumbents would rather acquire startups with de-risked technology and proven teams than fund that development internally.
Recent deals show this in practice. Marvell acquired Celestial AI, an optical interconnect startup, in a deal worth up to $5.5 billion, and XConn Technologies, a maker of interconnect silicon, for $550 million. Qualcomm acquired Alphawave Semi for $2.4 billion, Credo acquired DustPhotonics for up to $1.3 billion, and Nvidia paid $900 million to license Enfabrica’s networking technology and hire its team. In power delivery, Analog Devices acquired Empower Semiconductor, a Socratic portfolio company, for $1.5 billion[12]. None of these companies makes a processor. Each was bought because it removes a bottleneck that limits how fast or efficiently an AI system can run.
III. The Semiconductor Landscape Today: A New Strategic Investment Opportunity
The semiconductor industry is now experiencing substantial market expansion, unprecedented investment in AI infrastructure, and increasing strategic activity across technologies that were historically considered niche components of the broader semiconductor ecosystem.
Importantly, the current environment is not simply a continuation of the AI-driven GPU investment cycle. It represents a broader shift in the strategic importance of the technologies required to build, connect, and power increasingly complex computing systems.
A large and essential market, with or without AI
Before ChatGPT, semiconductors were one of the largest and most important industries in the world. In 2021, a year before generative AI reached consumers, the industry sold $556 billion of chips and shipped a record 1.15 trillion units[13]. Those chips went into nearly every product with a power source: phones, PCs, cars, factory equipment, medical devices, communications networks, satellites, and defense systems. The market has also grown steadily over time. Global sales rose from $139 billion in 2001 to $527 billion in 2023, a compound annual growth rate of about 6%, through the 2008 financial crisis and several inventory downturns[14]. The industry reinvests more than 15% of its sales in R&D each year, second only to pharmaceuticals and biotechnology among U.S. industries[15]. That is why technical innovation, rather than scale alone, decides who wins.
That growth was expected to continue without AI. In April 2022, seven months before ChatGPT launched, McKinsey projected the industry would reach roughly $1 trillion by 2030, growing 6% to 8% a year[16]. About 70% of that growth was expected to come from three markets: automotive, computation and data storage, and wireless. Automotive was the clearest example. As vehicles electrify and add driver assistance, McKinsey expected automotive chip demand to roughly triple, lifting its share of the market from 8% to as much as 15%. The shift was already under way: automotive chip sales grew 34% in 2021[17] and another 29% in 2022, reaching a record $34 billion[18].
The 2021 chip shortage showed how much of the broader economy depends on these products. Shortages cost automakers an estimated $210 billion in revenue and 7.7 million vehicles of lost production that year[19]. Governments responded by treating semiconductors as strategic infrastructure. The U.S. CHIPS and Science Act committed $52.7 billion to domestic manufacturing, research, and workforce programs, and the EU set out to mobilize €43 billion under its own Chips Act[20]. Before AI, then, the industry was already large, growing, diversified, and strategically important. As Section I describes, it was also underfunded by venture capital.
AI has added a new layer of demand on top of that foundation. In 2025 the market grew 26% to $796 billion, led by a more than 60% increase in computer segment sales tied to data center and AI investment[21]. WSTS now expects the market to reach about $1.51 trillion in 2026, passing the $1 trillion mark four years ahead of the pre-AI forecast[22]. Part of that jump reflects a sharp rise in memory prices. Even so, WSTS’s November 2025 forecast, which assumed far smaller memory gains, put 2026 at $975 billion[23]. Forecasters expect that growth to continue. WSTS projects about $1.9 trillion in 2027[24], and Gartner’s forecast is similar[25]. For 2030, Bank of America forecasts $3.2 trillion[26], and McKinsey’s base case is $2.3 trillion on a broader definition of the industry. Both are two to three times the pre-AI forecast for 2030. McKinsey attributes much of its increase to higher prices for leading-edge chips and memory, while still expecting wafer volumes to grow about 7% a year[27]
AI is also lifting demand beyond the data center processor. Amazon has said that growth in AI is driving growth in conventional computing, because reinforcement learning and agent tool use run mostly on CPUs rather than AI accelerators[28]. WSTS expects analog chips to grow 10% and discrete devices 8% in 2026, and these are categories that serve cars, factories, and power systems[29]. And as AI models move onto phones, PCs, vehicles, and robots, they raise the chip content of the same markets that drove the pre-AI forecast.
For investors, the case for semiconductors does not rest on AI alone. The core markets provide steady, diversified growth. AI adds a second engine on top of it and creates the new technical bottlenecks described in the following sections, which is where startups have the most room to compete.
Semiconductors are no longer a “niche” industry
Semiconductors are the foundation of the modern global economy, the way steel and steam powered the industrial revolution. Many of the world’s most valuable businesses are either semiconductor companies or depend on semiconductors to deliver their products and services. Three of the ten largest public companies by market cap, NVIDIA and Broadcom (fabless hardware vendors and designers), and TSMC (world’s largest fab), are semiconductor companies outright. Of the remaining seven, every company except one (Saudi Aramco) designs its own chips with an in-house silicon team. Apple builds its own A-series and M-series processors. Alphabet, Amazon, Microsoft, and Meta design custom AI silicon (TPU, Trainium, Maia, and MTIA) to improve the performance and economics of their AI workloads. SpaceX develops its own chips for Starlink. Model developers are now following the same path. OpenAI unveiled Jalapeño, its first custom AI chip, in June 2026, a processor built from scratch for large language model inference. In August 2026, Anthropic confirmed it is building an in-house silicon team to design custom chips for Claude.
None of these companies operate in isolation. Their products and roadmaps depend on an interconnected global ecosystem of chip designers, EDA and IP providers, semiconductor equipment manufacturers, memory suppliers, advanced packaging specialists, and foundries. That ecosystem extends to the buyers of chips as well. Hyperscalers and neoclouds purchase accelerators from chip suppliers and rent that compute to model developers. Capital flows back the other way: hyperscalers and chip suppliers invest in model developers, and chip suppliers invest in the neoclouds that buy their chips.
For startups, this web of relationships means a breakthrough anywhere in the stack has an unusually large pool of potential customers, partners, and acquirers, each with the motive and the capital to secure it.
Unprecedented demand for compute from training to inference to agents
The emergence of generative AI has fundamentally changed the semiconductor industry by creating a new and rapidly expanding source of demand for compute. The first phase of the AI buildout was dominated by training increasingly large models, driving demand for GPUs, accelerators, high bandwidth memory, networking, and data center infrastructure. The next phase is increasingly defined by inference, running these models continuously across billions of queries, applications, and autonomous workflows, and agentic AI, reasoning, using tools, executing multi-step tasks, and operating autonomously.
Agentic AI can require substantially more inference compute because a single task may trigger dozens or hundreds of model calls rather than one prompt and one response. Most markets are bound by population: a person needs only one car or one phone. Agents break that ceiling. A single user can run dozens of agents around the clock, each generating continuous inference demand, which makes compute consumption a function of work to be done rather than people to serve. Goldman Sachs estimates that agentic AI could drive a 24x increase in token consumption by 2030 as consumers and enterprises adopt these systems[30]. This sets up a self-reinforcing loop: better models lead to more useful applications, which drive more usage and more demand for compute, which funds more capable and efficient chips, which lower the cost per token and produce better models. AI systems that help build future AI systems (i.e., autonomous coding agents) could extend the loop further.
The scale of investment reflects the spending required to meet the expected demand. The five leading hyperscalers, Amazon, Google, Meta, Microsoft, and Oracle are expected to spend more than $800 billion in capex in 2026, based on public filings. This number is expected to increase nearly 4x over the next few years, reaching $3 trillion by 2030[31].
[32]
The hyperscalers are not the only ones allocating massive spending budgets to the buildout. Neoclouds (CoreWeave, Nebius, Lambda), data center developers (Equinix, Digital Reality, Crusoe), power providers (Constellation Energy, Vistra, NextEra), and the model developers (OpenAI, Anthropic, xAI) are all committing significant capital of their own. For example, OpenAI’s Stargate project is a $500 billion AI infrastructure plan whose costs are expected to be shared among OpenAI, Softbank, Oracle, and MGX. Anthropic’s S-1 filing shows that they expect to spend $518 billion in building out AI infrastructure over the next decade with six partners, including at least $111.1 billion with Google, $110 billion with Amazon, and $31.4 billion with Microsoft, plus $161.2 billion of Broadcom-related equipment lease obligations[33].
The semiconductor industry is the direct beneficiary of this infrastructure race. WSTS now projects the global semiconductor market to reach approximately $1.51 trillion in 2026, up roughly 90% year over year, and approaching $1.9 trillion in 2027[34].
Falling token prices expand the market for compute
As AI becomes a utility, the token has become its unit of sale, and the price of a token has become one of the most closely watched numbers in the market. Silicon Data’s LLM Token Expenditure Index, which tracks the average price paid for a million tokens across the market, fell by more than half between its May peak and the end of August, to roughly $0.97[35]. Some investors, setting that decline beside softer GPU rental prices over the summer, have read it as a sign that AI is losing its pricing power and that the infrastructure buildout will slow. The data points the other way.
The index measures price, not demand. Because it is weighted by usage, it falls whenever customers shift work to cheaper models, even if no list price changes, and that shift explains most of the 2026 decline. Open-weight models, which anyone can download and run, now handle most routine work: on Vercel’s AI Gateway they processed 56% of tokens in August but only 14% of spending[36]. Usage kept compounding underneath the falling price, with token volume on OpenRouter up roughly 47% in August alone[37]. Meanwhile, the price of compute moved in the opposite direction. The H100 rental index has risen more than 40% since December as on-demand capacity sold out[38].
Cheaper tokens are good news for the companies that build and operate compute. A token requires the same silicon, memory, and power to produce whether the model behind it is open or closed. What changes is who keeps the margin. Frontier labs earn high margins on their own models and buy compute at a scale that lets them negotiate hard. The thousands of companies serving open models cannot, which makes open weight tokens the highest margin tokens for the clouds and neoclouds[39]. As open models take share, margin moves from the model layer to the infrastructure layer. Cloud results already reflect this: in the second quarter, AWS revenue grew 37% to $42.2 billion at a 36.8% operating margin, and Google Cloud grew 82% while more than tripling its operating income[40]. Compute contracts signed in 2024 and 2025 are also rolling off into a market where spot rates sit well above them, lifting revenue on capacity already installed[41]. Those cash flows, not debt alone, fund the next round of capex.
Lower prices also expand demand. Each step down in price makes new applications economic, and agentic workflows, as described above, multiply the tokens consumed per task. Open-weight models have added to frontier usage rather than replaced it: their share of tokens on Vercel’s gateway rose from 28% to as high as 62% over two months this summer, even as combined usage of OpenAI and Anthropic models accelerated[42]. China offers a preview of the pattern. Daily token calls there rose from 100 billion in early 2024 to more than 140 trillion by March 2026, while H100 rental prices there rose 20% to 30% between October 2025 and April 2026[43].
Two conclusions follow for semiconductor investors. First, falling token prices are consistent with rising demand for chips: more tokens, each costing the same compute to produce, sold by infrastructure providers earning higher margins. Second, as the token becomes a commodity, cost per token becomes the measure buyers care about most. That cost is set less by the processor than by the system around it: how efficiently data moves, how much memory sits beside the chip, and how much power each token consumes. Those are the bottlenecks described in the next section, and they are where startups have the most room to win.
Compute cycle bottlenecks – more than just the GPU
The demand doesn’t stop at the processor. As described in earlier sections, training, inference, and agentic AI each require specialized AI processors, fast data movement through networking and optical interconnects, high bandwidth memory, efficient power delivery, and advanced packaging. As AI systems have grown from single GPUs to clusters of tens or even hundreds of thousands of accelerators, performance depends as much on the components that surround the chip as on the chip itself. No single processor can hold a frontier model or train it in a reasonable amount of time, so the work is split across many chips that must constantly share results. A training step cannot finish until every chip has exchanged its updates with the others, which means the cluster runs only as fast as its slowest connection. When data moves too slowly, expensive GPUs sit idle waiting for it.
The same is true once a model is deployed. For inference, every request must run across a group of accelerators that pass data back and forth as the model generates each token of its response. Memory needs also grow during use. A model keeps a running record of each conversation, known as the KV cache, and that record expands with every token of context.
As a result, the industry’s central engineering challenge has shifted from building a faster chip to making thousands of chips behave like one giant computer. Cluster sizes reflect this: models that were trained on a few thousand GPUs only a few years ago now run on clusters of 100,000 or more, spread across entire buildings and increasingly across multiple data centers.
Connecting these systems happens at two key levels. Scale-up links a group of accelerators so tightly that they function as a single, larger processor, sharing memory and exchanging data at extremely high bandwidth and low latency. NVIDIA’s GB200 NVL72, for example, connects 72 GPUs within one rack through its NVLink fabric. Today these links are mostly copper, but copper loses signal quality quickly over distance and at higher speeds. As scale-up domains expand from dozens to hundreds of accelerators across multiple racks, and the speed data travels increases from 100G, 200G, to 400G and beyond, the industry is moving toward optical connections, including co-packaged optics that place the optical engine directly beside the processor. Scale-out connects these racks to one another across the data center using Ethernet or InfiniBand networks, where pluggable optical transceivers carry data between switches over longer distances.
There is also scale-across, which links separate data centers when a single site can no longer supply enough power or space as well as scale-in, providing interconnect inside a single package or board, linking chiplets, compute dies and memory (such as HBM) at very short reach.
Each tier creates its own demand for semiconductors: switch chips, optical transceivers, lasers, photodetectors, silicon photonics, SerDes, and the advanced packaging needed to integrate optics alongside compute. Networking is also becoming a significant consumer of power in its own right, which makes energy per bit as important a metric as bandwidth. This is why optical interconnect is one of the fastest-growing segments of AI infrastructure. It is also where startups with novel approaches to bandwidth density, power efficiency, or cost, such as microVCSEL- or microLED-based links and new CPO architectures have a real opening to displace incumbent technology.
The critical investment implication is that AI does not simply create demand for more GPUs. It creates demand for an entirely new generation of semiconductor technologies required to support the broader AI infrastructure buildout.
Each of these components is a market in its own right, which significantly expands the addressable opportunity for semiconductor startups.
Incumbents are structurally poorly placed to move quickly in these key areas, yet innovation is required as soon as possible to solve the key bottleneck areas plaguing the industry today:
- GPU underutilization: The most expensive chips in the data center routinely spend much of their time idle, waiting on data. In large training runs, accelerators often deliver well under half of their theoretical peak performance. Every percentage point of utilization recovered is worth billions of dollars of capex across the industry, but the fix usually lies outside the GPU, in networking, memory, and software.
- The memory wall: Processor compute has grown far faster than the bandwidth of the memory that feeds it. AI models, especially inference with long contexts and agentic workloads, are increasingly limited by how fast data reaches the processor rather than by how fast it can be processed.
The interconnect wall: As clusters scale from racks to buildings, copper links run out of reach and power efficiency, and the network becomes the system’s speed limit. Optical interconnect is the clear path forward, but it requires new lasers, photonic integration, and packaging approaches that incumbents largely lack in-house. - The power wall: AI rack power has climbed from roughly 10–15kW to around 130kW today, and AI data centers are expected to need 1MW or more per rack before 2030. That requires a full redesign of power conversion, from the grid to the chip.
None of these problems can be solved by a better GPU alone, and solving any of them means changing an architecture that an incumbent has already committed to customers. This is a key opening for venture-backed innovation: a startup can design around a bottleneck from a clean sheet, taking risks that incumbents are electing not to.
Strategic exits for semiconductor companies are accelerating – and so is the path to liquidity for early investors
Incumbents already have their R&D budgets and engineering teams committed to publicly announced, multi-year product roadmaps, so building these capabilities internally is slow, expensive, and risky. A large chipmaker cannot easily bet a product generation on an unproven architecture, but a focused startup can. Acquiring a startup delivers both the technology and the team that knows how to keep developing it. In semiconductors, where design, verification, and manufacturing know-how is hard to replicate, the team can be worth as much to an acquirer as the technology itself.
The recent exit environment reflects this exact point. According to PitchBook, total semiconductor exits completed for VC-backed companies across M&A and IPO activity was just $15.1 billion in 2024, slightly above where it stood in 2020. It has quickly climbed from $22.0 billion in 2025 to $101.6 billion in 2026 as of the end of Q3, more than 4.5x the prior year with a quarter still to go.
The size of individual deals tells the same story. Semiconductor transactions of $1 billion or more have climbed from just one each in 2023 and 2024, to ten in 2025, with seven more announced so far in 2026. Many of these were venture-backed companies acquired well before reaching commercial scale, including Celestial AI, Enfabrica, Rivos, DustPhotonics, Empower Semiconductor, and Groq.
The pattern of recent exits and accelerating M&A activity shows that incumbents increasingly prefer to buy innovation rather than build it, and they are willing to pay significant premiums to do so. In the last 2 years, several companies have completed their largest deals in company history. Some examples include Skyworks’ acquisition of Qorvo for $22bn, Nvidia’s acquisition of Groq for $20bn, Texas Instruments’ acquisition of Silicon Labs for $8bn, Onsemi’s acquisition of Synaptics for $7bn, Arm’s acquisition of Ampere for $7bn, SiTime’s acquisition of Renesas’ timing business for $3bn, Lattice Semiconductor’s acquisition of Ami for $2bn, and Credo’s acquisition of DustPhotonics for $1.3bn.
While the price tag of recent exits may be staggering when compared to historical transaction values, the premiums paid are just another line item on the financial statements for the pool of acquirers. With market caps in the trillions and annual revenues in the multi hundred billions, buying a company in the double digit billion range is relatively insignificant. Nvidia alone generated $96.2 billion in revenue and $21.3 billion in free cash flow in their most recent quarter this year – the $20 billion Groq transaction was less than one quarter of cash flow[44].
The economics also favor buying over building. Developing a new capability internally means years of R&D spending that runs through the income statement as operating expense, weighing on the margins public investors watch most closely. An acquisition is paid for with balance sheet cash and delivers proven technology and an experienced team from day one. For an incumbent with tens of billions in cash and a shortage of specialized engineers, paying a premium for a derisked startup is often cheaper, faster, and less disruptive than building from scratch.
Private capital is flowing back into semiconductors
Investors are taking notice. Semiconductors are now attracting generalist VCs, growth investors, private equity, and corporate venture arms, a sharp change from a decade earlier. Private semiconductor funding has grown from roughly $1 billion in 2016 to $14.5 billion in 2026 as of September 30. That is already 49% above full-year 2025, with a quarter still to go. More than $36 billion has been raised since the start of 2024 alone.
Most of that capital has gone to the most visible part of the market: datacenter processing, led by companies like Groq, Cerebras, Tenstorrent, and Etched. While there remains significant value potential in the processor market, the market is starting to catch on to the equally important sectors that have flown “under the radar” such as connectivity, networking, and power which, as discussed previously, are significantly large markets on their own.
This creates an opportunity for investors with the technical expertise to look beyond the headline AI chip companies and identify the infrastructure bottlenecks that will define the next phase of the industry. The current flow of capital is an important signal that the market is beginning to recognize the opportunity, but the investment cycle is still broadening. As capital continues to move downstream from the processor into the broader semiconductor stack, some of the most attractive opportunities will emerge in the specialized technologies that enable the next generation of AI infrastructure.
Looking ahead: where the industry is headed
The first phase of the AI buildout was about scarcity: getting enough compute to exist, almost regardless of cost. The next phase is about efficiency. As AI moves from experiment to infrastructure, the metric that matters shifts from raw performance to performance per dollar, performance per watt, and tokens per second. Nvidia reports that its GB300 NVL72 rack lowers cost per token by up to 35x compared with its prior Hopper generation, a gain it attributes to designing the chips, networking, and software together[45]. Over the same period, AI is expanding outside the walls of the data center to phones, autonomous vehicles, factories, and robots, creating demand for chips that can run capable models on a fraction of the power. Each of these shifts is a new market driven forward by startup innovation.
The way chips are built is changing too. AI models are already helping engineers design and verify silicon, which could shorten development cycles and lower the cost of bringing a new chip to market. Governments are investing to bring manufacturing closer to home. Hyperscalers and model developers are turning into chip designers. The result is an industry with more builders, more buyers, and more problems worth solving than at any point in its history. For most of the last decade, semiconductors were treated as a mature industry. The years ahead will look much more like its early ones: fast-moving, inventive, and full of opportunity for the companies, and the investors, willing to build what comes next.
IV. The Semiconductor Investment Case
1. A large, essential market, with or without AI: Semiconductors grew about 6% a year for two decades and were on track for $1 trillion by 2030 before ChatGPT. AI has pulled that milestone forward to 2026, and 2030 forecasts now run from $2.3 to $3.2 trillion.
2. Innovation has moved to the system: With Moore’s Law slowing, performance gains now come from architecture, memory, interconnect, power, and packaging. Each is a large market where focused startups can move faster than incumbents tied to existing roadmaps.
3. Underinvested and under-competed: Semiconductor startups received less than 1% of U.S. venture capital from 2013 to 2022. Today’s funding still concentrates in AI processors, which leaves the bottleneck technologies open to specialist investors.
4. More capital efficient than assumed: The foundry model means startups spend on engineers and tape-outs, not fabs, and can prove their technology on shared wafer runs before committing to production.
5. AI demand is compounding, not peaking: Agentic AI could drive a 24x increase in token consumption by 2030, and falling token prices expand usage rather than shrink it. Every cheaper token still needs silicon, memory, and power.
6. Validated paths to liquidity: Exits of venture-backed semiconductor companies reached $101.6 billion through Q3 2026, up from $15.1 billion in 2024. Incumbents with tens of billions in cash are buying the technology and teams they cannot build fast enough.
Socratic Partners, November 2026
Disclaimer
This white paper has been prepared by Socratic Partners I LP (“Socratic ”) and is not an offer or solicitation and may not be used or relied upon in connection with any offer or solicitation, with respect to Socratic or any other future Socratic investment partnership. Investments in Socratic will involve significant risks, including risks of loss of the entire investment. Investors should have the financial ability and willingness to accept the risk characteristics of early stage, illiquid investments.
Some of the information presented herein may constitute “forward-looking information.” Forward-looking information is based on numerous assumptions and is speculative in nature and may vary significantly from actual results. Certain performance data contained herein may be based upon or include the value of unrealized investments. In many cases, the value of an unrealized investment is estimated due to the lack of a trading market. There can be no assurance that estimated values or returns based on those values can be realized or that actual returns or results will not be materially lower than those stated herein.
Certain assumptions have been made to simplify the presentation and, accordingly, actual results may differ materially from those presented. In considering any performance information contained herein, prospective investors should bear in mind that past performance is not necessarily indicative of future results, and there can be no assurance that Socratic will achieve comparable results in the future. There can be no assurance that Socratic ’s investment strategy will achieve profitable results.
The information in this white paper is intended to be current only as of the date of this presentation.
Footnotes
[1] Moore’s Law describes the historical trend of transistor density roughly doubling every two years, enabling continued improvements in computing performance and energy efficiency at declining cost per transistor.
[2] Bank of America Makes Massive Call on AI Chip Market
[3] Investor Relations | Marvell Technology, Inc. (MRVL)
[4] Fairchild Semiconductor: Origin of Silicon Valley – LIVE FROM SILICON VALLEY
[5] TSMC was founded in Hsinchu, Taiwan in 1987 by Morris Chang, a former Texas Instruments executive, with backing from the Taiwanese government and Philips. It was the world’s first dedicated foundry: it manufactures chips that other companies design and does not sell chips of its own. It is now the world’s largest foundry, with 72.5% of the revenue of the ten largest foundries in the second quarter of 2026, and in 2025 it made more than 12,000 products for 534 customers, including NVIDIA, Apple, AMD, and Broadcom. Sources: TSMC Company Profile; TrendForce, 2Q26 foundry rankings.
[6] PCAST Semiconductors-Report, Sept 2022
[7] PCAST Semiconductors-Report, Sept 2022
[8] Chipping Away: Assessing and addressing the labor market gap facing the US semiconductor industry | Oxford Economics
[9] How Many People Use ChatGPT? 2026 Statistics
[10] State of Markets II – by David George – a16z
[11] 2025: The State of Generative AI in the Enterprise | Menlo Ventures
[12] Deal values from PitchBook and company announcements. Values for Celestial AI and DustPhotonics include earnouts tied to future milestones.
[13] Global Semiconductor Sales, Units Shipped Reach All-Time Highs in 2021 | SIA
[14] 2024 SIA Factbook | Semiconductor Industry Association
[15] 2024 SIA Factbook | Semiconductor Industry Association
[16] The semiconductor decade: A trillion-dollar industry | McKinsey
[17] Global Semiconductor Sales, Units Shipped Reach All-Time Highs in 2021 | SIA
[18] Global Semiconductor Sales Increase 3.3% in 2022 Despite Second-Half Slowdown | SIA
[19] Chip shortage expected to cost auto industry $210 billion in 2021 | CNBC
[20] Chip Act versus Chip Act | Evertiq
[21] Global Semiconductor Market grows 26% in 2025 to $796B | WSTS
[22] Global semiconductor market surges beyond USD 1.5 trillion in 2026 | AnySilicon
[23] Global Semiconductor Market Approaches $1T in 2026 | WSTS
[24] Global semiconductor market surges beyond USD 1.5 trillion in 2026 | AnySilicon
[25] Gartner Forecasts Worldwide Semiconductor Revenue to Reach $1.6 Trillion in 2026 | Gartner
[26] Bank of America Makes Massive Call on AI Chip Market | Yahoo Finance
[27] McKinsey raises 2030 chip market forecast to $2.3T | Bits&Chips
[28] Amazon 2Q26: after Google and Microsoft, another cloud acceleration I Tech Stock
[29] Global semiconductor market surges beyond USD 1.5 trillion in 2026 I AnySilicon
[30] AI Agents Forecast to Boost Tech Cash Flow as Usage Soars I Goldman Sachs
[31] Investor Relations I Marvell Technology, Inc (MRVL)
[32] Source: Company SEC Filings.
[33] Anthropic’s $518 billion AI buildout hinges largely on deals that cannot be cancelled, filing shows I Reuters
[34] Global Semiconductor Market Surges Beyond $1.5T 2026
[35] AI Token Price Collapse: Goldman Sachs Warns the Unit Economics of the AI Boom Are Breaking I Tiger Brokers
[36] AI Gateway Production Index, September 2026 I Vercel
[37] AI Token Price Collapse: Goldman Sachs Warns the Unit Economics of the AI Boom Are Breaking I Tiger Brokers
[38] H100 Rental price Index I Silicon Data
[39] Plot Twist: The ‘Highest Margin Tokens’ for Amazon, Microsoft and CoreWeave are Open-Weight Models I TIKR
[40] Cloud big three post record growth despite AI bubble fears I Digital Today
[41] Gavin Baker on X: hyperscalers are underearning as compute reprices (July 2026)
[42] Gavin Baker on X: hyperscalers are underearning as compute reprices (July 2026)
[43] NVIDA H100 Rentals in China Reportedly Up 20%-30% as Token Demand Surges I TrendForce
[44] NVIDIA (NVDA) Q2 2027 Summary | Quartr
[45] What drives cost per inference request at scale I NVIDIA