The numbers have become difficult to process with a straight face. OpenAI closed a secondary share sale in March 2026 that valued the company at $350 billion, roughly 35 times its annualized revenue run rate of $10.2 billion. Anthropic’s latest funding round in January pegged the Claude developer at $165 billion, up from $61 billion just fourteen months earlier. Mistral, the Paris-based foundation model builder that didn’t exist three years ago, is reportedly seeking a valuation north of $40 billion for its Series C.

Across the broader AI startup ecosystem, PitchBook data shows that venture-backed AI companies raised $128 billion in 2025, more than double the $56 billion raised in 2023. The median pre-revenue AI startup is now raising its Series A at a $180 million valuation, a figure that would have been considered absurd for a growth-stage company a decade ago. The question investors, analysts, and founders are now wrestling with is whether these numbers represent the rational pricing of a generational technology shift or whether the AI funding cycle has entered a phase that will be studied in future business school case studies alongside Pets.com and WeWork.

The honest answer is probably both.

The Case for Bubble: Revenue Multiples That Defy Historical Norms

By virtually any traditional valuation metric, the AI sector is in uncharted territory. The median revenue multiple for public SaaS companies sits at roughly 7x forward revenue as of April 2026, according to Bessemer Venture Partners’ Cloud Index. The top-tier foundation model companies are trading at 25x to 40x revenue in private markets. Even during the peak of the 2021 ZIRP-fueled software bubble, the most aggressively valued public cloud companies rarely exceeded 50x revenue, and those multiples collapsed by 70% to 80% in the correction that followed.

The comparison to the dot-com era is instructive but imperfect. In 1999, the Nasdaq traded at roughly 175 times earnings. Companies with no revenue and no clear path to profitability commanded multi-billion-dollar market capitalizations. Today’s AI startups are not, for the most part, revenue-free. OpenAI genuinely generates over $10 billion in annual recurring revenue. Anthropic has crossed $3 billion ARR. These are real businesses with real customers paying real money. But the gap between current revenue and the valuations being assigned suggests investors are pricing in a future where these companies capture a share of global GDP that would be historically unprecedented for any technology sector.

Goldman Sachs published a research note in February 2026 estimating that the AI infrastructure buildout — including data centers, chips, energy, and model training — will require roughly $1.5 trillion in cumulative capital expenditure by 2030. The report’s central finding was sobering: for AI investments to generate adequate returns, the technology needs to generate approximately $600 billion in incremental enterprise revenue by 2030, a figure that implies a roughly 20x increase from current levels. Goldman’s analysts described the gap between current AI revenue and required future revenue as “the trillion-dollar question mark.”

The Sectors Most at Risk

Not all AI valuations carry equal risk. The frothiest segment of the market is the application layer — companies building AI-powered tools for specific verticals like legal, healthcare, customer support, and marketing. Many of these startups have raised $50 million to $200 million at valuations of $500 million to $2 billion, despite having annual recurring revenue in the single-digit millions and gross margins that are structurally compressed by the cost of API calls to foundation model providers.

The problem is straightforward: application-layer AI companies are, in many cases, thin wrappers around foundation model APIs. Their differentiation rests on prompt engineering, fine-tuning, and domain-specific data — all of which are increasingly commoditized as foundation models improve. When GPT-5 or Claude 4.5 can perform legal document review or medical coding at near-expert level out of the box, the value of a startup that built its entire product on doing those tasks with GPT-4 diminishes rapidly.

Sequoia Capital’s internal analysis, portions of which leaked to The Information in March 2026, estimated that roughly 60% of AI application-layer startups funded in 2024 and 2025 will fail to reach their next funding milestone. The memo reportedly used the phrase “the AI wrapper reckoning” to describe the coming shakeout.

The Case for Boom: Where Real Revenue and Durable Moats Exist

The bubble narrative, however, misses something important: the foundation model layer is producing genuine, rapidly growing revenue that bears no resemblance to the vapor economics of 1999.

OpenAI’s trajectory is the clearest evidence. The company grew from $3.4 billion in ARR at the end of 2024 to $10.2 billion by March 2026, a tripling in roughly fifteen months. Enterprise adoption is the primary driver. Microsoft’s integration of OpenAI models into its Office suite, Azure cloud platform, and GitHub Copilot has created a distribution channel that touches hundreds of millions of users. Roughly 85% of Fortune 500 companies now use at least one OpenAI-powered product, according to Microsoft’s most recent earnings disclosure.

Anthropic’s growth has been equally striking, if from a smaller base. The company’s Claude models have become the preferred choice for enterprises prioritizing safety, reliability, and long-context performance. Amazon Web Services, which has invested more than $8 billion in Anthropic, has made Claude the flagship model on its Bedrock platform. Anthropic’s enterprise revenue grew from roughly $800 million in ARR at the end of 2024 to $3.2 billion by early 2026, driven heavily by financial services, healthcare, and government clients.

The Infrastructure Layer Is Printing Money

The clearest winners in the AI economy are not the model builders — they’re the companies selling the picks and shovels. Nvidia reported $38.2 billion in revenue for its most recent quarter, with data center GPU sales accounting for more than 80% of the total. The company’s gross margins exceed 75%, a figure that would be considered extraordinary in any hardware business. Taiwan Semiconductor Manufacturing Company (TSMC), which fabricates the vast majority of advanced AI chips, has seen its revenue grow 35% year-over-year, with AI-related orders now representing more than 40% of its advanced node capacity.

Cloud infrastructure providers are similarly benefiting. Amazon Web Services, Microsoft Azure, and Google Cloud collectively reported more than $80 billion in AI-related cloud revenue in their most recent fiscal years, with AI workloads growing at roughly 60% annually. These are not speculative revenue streams. They represent real computing consumption by real enterprises running real AI workloads.

The Competitive Landscape: A Three-Way Race With Wild Cards

The foundation model market has consolidated faster than many predicted. OpenAI, Anthropic, and Google DeepMind represent the clear top tier, with Meta’s LLaMA open-source models and Mistral occupying strategically important but commercially distinct positions.

OpenAI’s advantage is distribution. Through its partnership with Microsoft, OpenAI’s models are embedded in the productivity tools used by over a billion people. The company’s consumer products — ChatGPT and its successors — have over 400 million monthly active users. This distribution advantage creates a data flywheel that is difficult for competitors to replicate.

Anthropic’s advantage is trust. The company has positioned itself as the safety-focused alternative, and that positioning has resonated powerfully with regulated industries. Banks, insurance companies, healthcare systems, and government agencies — the sectors most sensitive to AI risk — have disproportionately chosen Anthropic’s models. Claude’s performance on enterprise reliability benchmarks and its Constitutional AI framework have given the company a differentiated position that goes beyond raw model capability.

Mistral represents the most interesting wild card. The French company has positioned itself as the European alternative, benefiting from EU data sovereignty concerns and the European Commission’s stated preference for supporting homegrown AI champions. Mistral’s latest models have achieved performance parity with GPT-4-class models at significantly lower inference costs, making them attractive to cost-sensitive enterprises.

The Open-Source Disruption Risk

The most significant long-term risk to foundation model valuations may come from open-source models. Meta’s LLaMA 4 models, released in early 2026, have achieved performance levels that rival proprietary offerings on many benchmarks. The implication is uncomfortable for investors paying 35x revenue for OpenAI: if open-source models continue closing the capability gap, the pricing power of proprietary model providers could erode significantly.

DeepSeek, the Chinese AI lab backed by the quantitative trading firm High-Flyer, has demonstrated that competitive foundation models can be built for a fraction of the cost assumed by Western labs. DeepSeek’s R1 model achieved state-of-the-art performance on several reasoning benchmarks while reportedly being trained on a budget of less than $10 million, compared to the hundreds of millions spent by Western competitors on comparable models.

What History Suggests — And Where It Falls Short

Every technology bubble in history has contained a core of genuine innovation surrounded by a periphery of speculative excess. The dot-com bubble destroyed hundreds of billions in investor wealth, but it also produced Amazon, Google, and the modern internet economy. The crypto bubble of 2021 wiped out trillions, but it also produced the stablecoin infrastructure and decentralized finance protocols that are now being integrated into traditional banking.

The AI cycle appears likely to follow a similar pattern. The foundation model companies with real revenue, genuine enterprise adoption, and defensible competitive positions — OpenAI, Anthropic, and perhaps two or three others — will likely justify their valuations over a five-to-ten-year horizon, even if they experience significant multiple compression in the interim. The hundreds of application-layer startups with thin moats, compressed margins, and valuations built on narrative rather than revenue will face a reckoning.

Morgan Stanley’s technology team published a framework in March 2026 for evaluating AI investments that distilled the question to a single metric: gross margin sustainability. Companies that own their model infrastructure and can maintain gross margins above 60% as they scale are likely building durable businesses. Companies that are reselling foundation model API calls at a markup and operating at 30% to 40% gross margins are vulnerable to both margin compression and disintermediation.

The AI startup valuation landscape in 2026 is neither purely bubble nor purely boom. It is a market in which transformative technology and speculative excess coexist in uncomfortable proximity. The trillion-dollar question is not whether AI will reshape the global economy — at this point, that outcome appears virtually certain. The question is how much of the value creation will accrue to the current crop of venture-backed startups versus the incumbents, open-source alternatives, and companies that haven’t been founded yet. History suggests the answer will surprise nearly everyone betting on it today.