On May 20, 2026, three things collided on the same day.
SpaceX publicly filed its S-1 prospectus with the SEC, targeting a $1.75 trillion valuation and a June 12 Nasdaq listing. The Wall Street Journal reported exclusively that OpenAI could file its confidential IPO prospectus as early as Friday, aiming for September. And Anthropic’s annualized revenue had already surpassed OpenAI’s — Axios noted in its April column the $30 billion to $25 billion figure, while Counterpoint Research data confirmed Anthropic’s 31.4% global LLM revenue share, putting it in first place.
The combined valuation of these three companies exceeds $3.5 trillion. This will be the largest concentration of technology company IPOs in public market history. The dominant media narrative focuses on “who drains the capital pool first,” but this framing misses a deeper question: these three prospectuses aren’t the endpoint of a three-way company competition — they are three answers the entire AI industry is submitting to public markets. The question isn’t “how much can you earn.” It’s “what are you betting the endgame of AI looks like.”
The core conviction of the AI industry over the past three years can be summed up in one sentence: more compute equals better models. This belief powered every leap from GPT-3 to GPT-5 and justified each successive record-breaking funding round. But entering 2026, the evidence is pointing in a different direction.
On standard benchmarks, the gap between models is collapsing. Stanford HAI’s 2026 AI Index shows that the top 15 models on MMLU-Pro are clustered within less than five percentage points of each other. The independent benchmarking platform LLM Stats found that 8 of 14 major benchmarks are now “saturated” — top-performing models are approaching ceiling performance. Kilo AI ran a direct comparison: GPT-5.1, Gemini 3.0, and Claude Opus 4.5 were given the same real-world coding tasks, and the results were nearly identical.
Meanwhile, inference costs are dropping 10x per year. GPT-4-level performance cost $60 per million tokens at the end of 2022. By the end of 2025, it had fallen to $0.40. a16z’s analysis puts it bluntly: “intelligence is becoming too cheap to meter.” If this trend continues, the logic of building a moat solely on “I have a better model” becomes increasingly difficult to sustain.
Taken together, these signals point to an industry fracturing: from a single Scaling Law curve, splitting into three concurrent tracks — pretraining remains important but is no longer the sole driver; RL post-training and inference-time compute are taking over the growth baton.
Ilya Sutskever, after leaving OpenAI, publicly stated a line that has been widely quoted: “The 2010s were the age of scaling. Now we’re back in the age of wonder and discovery.” Sebastian Raschka, writing for Ahead of AI, systematically mapped the branches of inference-time compute research and cited a February paper showing that a 1-billion-parameter model with inference-time compute can outperform a 405-billion-parameter model without it.
This paradigm shift has left distinctly different traces on the capital expenditure of the three companies.
OpenAI’s Stargate project is nominally a $500 billion investment, but only $52 billion in equity has been committed, leaving a $448 billion gap to be filled by debt or new financing. The timing is what makes this critical: this investment lands precisely at the inflection point where the industry is transitioning from pretraining Scaling to a multi-track paradigm. If inference-time compute and RLVR become the primary engines of future capability leaps, then Stargate — this generation of compute centers built to “make models bigger” — could become the industry’s first “stranded assets.” This isn’t a bet on whether compute will be sufficient. It’s a bet on whether the physical path of capability improvement has been correctly identified.
Andrej Karpathy’s choice provides a precise footnote to this bet. On May 19 — the day after Musk lost his lawsuit — Karpathy announced he was joining Anthropic’s pretraining team. His specific mandate is to build a new team that uses Claude to automate Anthropic’s own pretraining research. This is a recursive proposition: use AI to build better tools for building AI. In his year-end review at the close of 2025, he wrote explicitly that RLVR — Reinforcement Learning from Verifiable Rewards — was the technical path that drove the largest capability leaps in 2025, and that labs are reallocating compute from pretraining to RL optimization.
He isn’t saying “pretraining is dead.” He’s saying pretraining needs to be reimagined — from stacking more GPUs and more internet-scale corpora, toward AI-assisted data curation, automated research, and efficiency optimization. His choice of Anthropic over OpenAI suggests he believes Anthropic’s investment in this new methodology is more thorough.
Anthropic’s revenue growth is the fastest in American corporate history. It went from $9 billion in annualized revenue at the end of 2025 to $30 billion in four months. Management’s updated guidance now suggests profitability could arrive in 2027 rather than 2028. The Information’s analysis suggests annualized revenue could reach $100 billion this year.
Much of this growth comes from deep enterprise integration. KPMG and Anthropic announced a global partnership: Claude is embedded in KPMG’s Digital Gateway platform, accessible to all 276,000 employees. PwC is certifying 30,000 professionals on Claude. Goldman Sachs engineers are co-developing AI agents with Anthropic for trade reconciliation, client onboarding, and compliance workflows — not a pilot, but embedded inside the bank’s operational infrastructure. ServiceNow made Claude the default model for its Build Agent, an agent that runs on ServiceNow’s platform processing 80 billion workflows annually. Over 1,000 enterprise customers spend more than $1 million annually on Claude — a number that doubled in two months.
These numbers tell a story of enterprise lock-in — an attempt to replicate the kind of deep integration that makes Salesforce or SAP “once you’re wired in, you can’t pull out.” Anthropic deploys simultaneously on AWS, Google Cloud, and Microsoft Azure, training on a mix of Trainium, TPU, and Nvidia GPUs. The multi-platform strategy itself sends a signal to enterprise customers: you won’t be locked into a single vendor’s ecosystem.
But there’s a critical counterexample.
a16z documented in an analysis that as agentic workflows proliferate, enterprises are indeed investing heavily in prompt engineering and guardrails, making them more reluctant to switch model providers. Yet the same analysis also recorded real feedback from one enterprise: “Switching models is a significant engineering effort, but precisely because of that, we designed flexible multi-model architectures from the start.”
ServiceNow is both Anthropic’s flagship customer and the embodiment of this paradox. It made Claude the default model for its Build Agent — but simultaneously maintains a partnership with OpenAI for speech-to-speech and general automation. Snowflake signed a $200 million Anthropic deal without abandoning Google’s Gemini. Large enterprises are consciously avoiding betting on a single vendor.
This creates a tension. Anthropic’s revenues are soaring, customers are integrating deeply, but the underlying architecture is deliberately designed to be replaceable. This isn’t the hard lock-in of SAP — “you can’t replace my database.” This is soft lock-in — switching costs are real but not irreversible the way traditional enterprise software is. And the reality of inference costs dropping 10x per year means a competitor can always offer “half your price for roughly your performance” to break that lock.
The three companies occupy different positions on this bet. Anthropic is betting that deep enterprise integration can withstand model commoditization. OpenAI is betting on consumer brand and distribution — ChatGPT has 900 million weekly active users — and is using its ad beta to attempt to convert attention into predictable revenue streams. SpaceX is betting on the scarcity of the physical infrastructure layer — GPU clusters, orbital compute, launch capability — wagering that no matter whose model you use, you have to run on someone’s GPUs.
This bet is the most asymmetric among the three, and the most easily misread.
Anthropic has consistently positioned itself as “the safest AI lab.” Its corporate structure is a Public Benefit Corporation. Its brand narrative revolves around constitutional AI alignment and responsible scaling policies. This narrative has commanded a premium in private markets — investors are willing to pay higher valuation multiples for “safe AI.”
But in March, something happened that directly challenged this premise. The Pentagon officially designated Anthropic a “supply chain risk” under 10 U.S.C. § 3252, barring defense contractors from commercial dealings with Anthropic. The trigger: during contract negotiations, Anthropic refused to allow Claude to be used for fully autonomous weapons and mass domestic surveillance. Anthropic sued the Trump administration. Just Security’s legal analysis concluded that the Secretary of Defense’s action “vastly exceeded his statutory authority.”
Meanwhile, xAI’s Grok is being integrated into the Pentagon’s classified networks, with the Secretary of Defense publicly stating that this means “the ethical handcuffs are off.”
This is not a story about whether “defense contractor status confers regulatory immunity.” The evidence doesn’t support that proposition — the FTC sued Lockheed Martin in 2022 just the same. The regulatory leniency that SpaceX-affiliated companies have received operates through a different mechanism: Musk’s personal political alliance with the president. During Musk’s tenure leading the Department of Government Efficiency, the DOJ dropped its employment discrimination lawsuit against SpaceX, and multiple inspectors general at federal agencies investigating Musk’s companies were fired. This protection is non-transferable — it attaches to the person, not the institution. Anthropic paying SpaceX for compute does not mean Anthropic enjoys the same political protection. The facts point in the opposite direction.
The way these three companies are betting on the same compliance question reveals their different understandings of what enterprise customers need. Anthropic is betting on institutional compliance — accepting government regulatory constraints, winning enterprise customers (banks, law firms, hospitals) that are themselves regulated, under the brand of “safe and responsible.” OpenAI is betting on consumer-side regulatory exemption — products aimed at individuals face far less compliance pressure than products aimed at enterprises. SpaceX is betting on political capture — relying on the founder’s personal network to shape the regulatory environment rather than on a compliance department.
These three strategies can coexist in private markets. But once they enter public markets, the risk factors section of each prospectus will force each company to explain to investors: does your compliance strategy still hold after a change in administration? How much premium remains in Anthropic’s safety brand under a Pentagon blacklist? Will OpenAI’s advertising monetization trigger cascading state-level privacy legislation? If SpaceX’s regulatory advantage evaporates with the political cycle, how many of those $22 billion in defense contracts are genuinely irreplaceable? Right now, 1,561 AI-related bills sit in 45 state legislatures, and the federal “light touch” framework could be rewritten the moment political power changes hands.
The true weight of these three prospectuses isn’t about which of the three companies is stronger. Their weight lies in the fact that public markets are being asked, for the first time, to put a price on three questions that have no consensus answers.
The first question: which path will AI’s capability leaps follow next? If you believe bigger pretraining is the only answer, then OpenAI’s Stargate is what you should buy. If you believe inference-time compute and RLVR will take over, then Anthropic’s pretraining efficiency strategy and Karpathy’s presence are more attractive. If you believe the physical scarcity of compute itself is the ultimate bottleneck, then SpaceX’s orbital data centers and GPU clusters are the most valuable.
The second question: to what extent can enterprise customers be locked in? If you believe AI switching costs can reach SAP levels — once wired in, impossible to pull out — then Anthropic’s 1,000 million-dollar customers are a replication of enterprise software economics, and a $900B valuation is justified. If you believe model commoditization and multi-model strategies will keep switching costs low, then the price elasticity of these enterprise contracts is greater than you think.
The third question: is AI’s compliance premium institutional or personal? If you believe Anthropic’s “safe AI” brand commands durable, irreplaceable pricing power over the long term, then the current regulatory environment is short-term noise. If you believe this industry’s regulatory outcomes depend on who holds power in Washington rather than who does better safety research, then Anthropic’s premium is a political derivative.
These three questions have no settled answers. But once the prospectuses are filed, the market must put a price on them. This is the real meaning of these three companies crowding into public markets in the same summer — not that they’re scrambling for capital, but that they’re being judged by the market.
Research Note: This article is based on publicly available information as of May 21, 2026. All specific figures and events are drawn from company SEC filings, official statements, the Wall Street Journal, CNBC, The New York Times, Axios, The Information, Bloomberg, Stanford HAI, a16z, and other independent sources. Views expressed represent analysis and judgment based on publicly available information.