OpenAI confirmed on June 8, 2026 that it had confidentially submitted a draft S-1 to the SEC. Anthropic did the same a week earlier, on June 1. SpaceX filed in May. Three of the four names that own the AI compute supply chain — three of the names whose decisions you have been building your agent stack around for the last two years — are now in the IPO queue. Reuters reports OpenAI is targeting up to a $1 trillion valuation, with a public offering potentially as early as September 2026. The private-market valuation sits at $730B after the most recent tender; Anthropic closed its last round at $852B; SpaceX is the SpaceX-sized number that nobody on Wall Street will say out loud.
The press is going to frame this as "AI is going public." That framing is wrong, and it is going to cost you money if you are building on top of the closed API tier. The correct framing is: **the closed AI labs just chose quarterly earnings pressure over the agent race, and that single decision reprices every build-vs-buy call you are going to make for the next eighteen months.**
I have been waiting for this. Not because I wanted it — because the trajectory has been obvious for at least nine months. Closed labs burning $5B to $13B a quarter on capex cannot stay private forever. The capex is the point. The IPO is the mechanism. What I did not expect is that it would land in the same week that NVIDIA shipped Nemotron 3 Ultra, in the same month that vLLM Semantic Router v0.3 Themis made open-weights routing a first-class problem, and in the same quarter that the agent stack consolidated around three or four serious runtime choices. The closed labs and the open stack are colliding on the same calendar. That is not a coincidence. It is the actual story.
Look at what the S-1 is going to have to disclose, and look at the contracts that are already public. The Google–SpaceX deal: $920M per month for 32 months from Google alone. The Anthropic–SpaceX deal: $1.25B per month through 2029. The Anthropic–Google–Broadcom multi-gigawatt TPU agreement. The Stargate commitments. The CoreWeave debt. The Oracle cloud build-out. The Microsoft carry-over from the 2023 agreement. NVIDIA's own data-center capex. None of this is on the income statement as "compute" — it shows up as "cost of revenue," "infrastructure commitments," and "related-party transactions," and the auditor is going to have a field day.
But here is the thing. The capex commitments are not a cost. **The capex commitments are the product.** OpenAI and Anthropic are not selling tokens. They are selling the option to be the inference tier for the next decade of agentic AI. The S-1 is not a financial document. It is a procurement contract with public-equity accountability attached. Once the S-1 is public, every quarterly call is going to answer one question: did you spend enough on capex to keep the closed API tier ahead of the open-weights tier? If the answer is no, the stock trades like a commodity. If the answer is yes, the stock trades like AWS in 2015.
This is the part most people miss. The closed labs are not optimizing for profit in 2026 or 2027. They are optimizing for the moment in 2028 or 2029 when the open-weights tier catches up on intelligence but cannot catch up on infrastructure. The capex now is the moat later. The IPO is the mechanism that funds the capex without giving the capex away to a sovereign wealth fund or a hyperscaler.
If you are building an agent in 2026, you are making a build-vs-buy decision every time you choose a model, a framework, a router, a memory layer, an eval harness, an observability stack, or an inference provider. Almost every choice you make is a bet on which API tier wins. Until June 8, that bet was a bet on private-market dynamics and founder preferences. After June 8, **that bet is a bet on quarterly earnings.**
Read that again. The closed API tier is no longer priced like a startup. It is priced like a public utility. The price of GPT-5.x, the price of Claude Opus 4.8, the price of Gemini 2.5, the price of the next three generations — all of those prices are now set against a $1T market cap and a board that needs to grow revenue 40% to 60% year-over-year to justify the multiple. That is not a price war. That is a price discipline regime. Token prices do not collapse. Token prices drift down 15% to 25% a year while the closed labs keep gross margins above 70% and use the cash to fund the next capex cycle.
If your agent is on a closed API, you are not buying compute. **You are buying a regulated utility contract with a public-equity counterparty.** That is a different risk profile. Regulated utilities do not collapse on price. They collapse on service. The agent that worked on GPT-5.x in Q3 2026 has to keep working on GPT-6 in Q1 2027 and on whatever the lab ships in Q3 2027. The model deprecates. The behavior changes. The eval you wrote in June 2026 fails in December 2026 because the model is now too capable in a way that breaks the guardrails you built around its old failure modes. You are paying for a moving target.
The open-weights tier is the only place where you can pin a model, pin a version, run it for two years, and not have the API tier shift under you. Nemotron 3 Ultra's NVFP4 checkpoint is the same artifact on Hopper, Blackwell, and Ampere. The license is the OpenMDW-1.1. The model card is the model. The community builds the inference stack. You are buying a depreciating capital asset, not a public-equity counterparty. Different risk, different reward, different bill.
I have been writing about this for a week. Nemotron 3 Ultra is the first open-weights release where the open lab has a defensible business model for staying open. NVIDIA does not need to make money on the model — NVIDIA makes money on the GPUs that run the model. The Nemotron family is the proof-of-concept for the next two years of inference hardware demand. Every team that runs Nemotron 3 Ultra on Blackwell is a customer for the next generation of NVIDIA hardware. The open-weights stack is not an alternative to the closed labs. It is the demand-generation arm of the closed labs' biggest supplier.
This is the part that changes the math. The open-weights tier is no longer the underdog competing with the closed labs on capital. The open-weights tier is the bet that the closed labs' biggest supplier is making on the long-term viability of inference as a hardware business. NVIDIA, AMD, Broadcom, the custom silicon shops — they all benefit from an open-weights tier that runs on commodity hardware and pushes the closed labs to spend more on the next-generation accelerators. The closed labs benefit because the open-weights tier keeps the developer ecosystem engaged, which keeps the model market liquid, which keeps API consumption growing. The open-weights tier is not the enemy of the closed labs. It is the second product of the closed labs' biggest supplier.
For you, the agent builder, that means the open-weights stack is not a fallback. It is the hedge. You run 60% of your traffic on closed APIs because the closed labs still have the best raw intelligence (Opus 4.8 at 61 on the AAII, GPT-5.x at 58, Gemini 2.5 Pro at 55). You run 30% of your traffic on open weights because the open stack is now fast enough, cheap enough, and good enough for the workload (Nemotron 3 Ultra at 48, Kimi K2.6 at 54, GLM 5.1 at 47). You run 10% of your traffic on a self-hosted distillate because the model pinning is the only way to guarantee behavior. That mix is not a compromise. That is the production stack for 2026 and 2027.
Here is the other thing the S-1 is going to do. It is going to clean up the agent framework market.
The closed labs are going to use the IPO proceeds to fund integration. OpenAI is going to ship first-party agent tooling that competes with LangGraph, CrewAI, AutoGen, and the rest of the framework stack. Anthropic already has Claude Code. Google already has ADK. The framework market is going to be compressed into three or four serious choices: the lab-native first-party stack for the closed API, an open-source framework that runs on every model, and a domain-specific framework that owns a vertical (Harvey for legal, Glean for enterprise search, Devin/Cursor for coding, Hippocratic for clinical).
If you are building a horizontal agent framework right now, the S-1 is the countdown clock. You have twelve to eighteen months to find a vertical, find a customer segment, and find a renewal. After that, the closed labs will have shipped the integration you were building, the open-weights tier will have shipped the cross-model abstraction you were building, and the only thing left for you to own is the domain expertise you have not yet built.
This is not a doom call. This is a call to focus. The horizontal agent framework is going to be the next-generation chatbot. The vertical agent framework is going to be the next-generation vertical SaaS. Pick a vertical. Build the deepest possible integration with the domain. Own the eval harness, the failure modes, and the customer trust. The S-1 just made that choice urgent.
I am not going to pretend I have read the S-1. It is confidential. But here is what I expect to see, based on the public statements from the labs and the comparable filings.
**Revenue growth in the 200% to 400% year-over-year range.** OpenAI and Anthropic are both growing faster than any public software company in history. The S-1 is going to show that. The multiple is going to be justified on growth, not on margin.
**Compute cost of revenue in the 35% to 55% range.** This is the line item that is going to spook the sell-side analysts. It is going to be the line item that anchors the comparison to AWS at a similar stage. The bull case is that the cost of revenue ratio falls over time as the labs negotiate better deals and ship more efficient models. The bear case is that the cost of revenue ratio stays high because the capex is the moat.
**Capex commitments in the $50B to $200B range over five years.** This is the number that is going to drive the stock. The S-1 is going to disclose contractual commitments to NVIDIA, AMD, Broadcom, Google, Amazon, Microsoft, Oracle, and CoreWeave. The number is going to be bigger than what the bears expect and smaller than what the bulls want. The stock is going to trade on the next quarterly update to that number.
**Related-party transactions.** Microsoft, Google, Amazon — they are all going to show up as both customers and suppliers. The S-1 is going to have to disclose those relationships. The audit committee is going to have a lot of work to do.
**The risk factors.** The risk factors are going to be the most interesting part of the S-1. They are going to name the open-weights tier as a competitive risk. They are going to name the regulatory environment as a competitive risk. They are going to name the talent market as a competitive risk. They are going to name the agent framework market as a competitive risk. The S-1 is going to be a public, legal description of why the closed labs are scared of the next eighteen months, and it is going to be the most-read document in tech.
If you are an agent founder, this is your week. Three things, in order.
**One: re-price your inference bill.** If you are 100% on closed APIs, model out the next twelve months at current prices. Model out the next twelve months at the 15% to 25% annual price decline that the labs are going to enforce. Model out the next twelve months at the 40% to 60% revenue growth the labs are going to report. The intersection of those three lines is your actual cost trajectory. Plan around it.
**Two: ship the open-weights fallback.** Nemotron 3 Ultra is fast enough for production. vLLM Semantic Router v0.3 Themis is stable enough for production. The SGLang, vLLM, and TRT-LLM day-zero support is real. You do not need to migrate all of your traffic. You need to migrate the 30% of your traffic that is the most cost-sensitive and the most behavior-sensitive. The 10% that needs the absolute frontier stays on closed APIs. The 60% that is routine agent execution moves to open weights. You will not get there in a week, but you will get the migration plan in a week.
**Three: pick your vertical.** If you have been building a horizontal agent, you have ninety days to pick a vertical and start building domain expertise. The S-1 is going to make the horizontal agent market a closed-lab-and-open-weights duopoly. The verticals are the only place where independent frameworks can survive. Pick the vertical you know. Pick the customer segment you can sell to. Pick the eval harness you can defend. Build the deepest possible integration. Own the renewal.
OpenAI's S-1 is not a financial event. It is a procurement event, a competitive event, and a repricing event. The closed AI labs just chose Wall Street discipline over the agent race, and that means every agent build-vs-buy decision for the next eighteen months is going to be a bet on public-equity dynamics. The open-weights tier is no longer the alternative. It is the only rational hedge. The agent framework market is going to compress into three or four serious choices, and the verticals are the only place where independent frameworks can survive.
The closed labs are still ahead on raw intelligence. The open-weights tier is still behind on absolute score. But the gap just closed, the throughput lead is real, and the business model behind the open release is the one the closed labs do not have. The agent stack for 2026 and 2027 is closed APIs for the frontier, open weights for the workload, and a self-hosted distillate for the behavior. The S-1 is not the end of the closed API tier. It is the moment the closed API tier became a regulated utility, and the moment the rest of the stack had to become a real hedge.
If you are building on the closed API tier, this is your warning. Reprice the bill. Ship the fallback. Pick the vertical. The S-1 is going to be public in a matter of months, and when it is, the market is going to reprice the closed labs in real time. You want to be on the right side of that repricing. You do not want to be the agent founder who bet the company on a quarterly earnings call.
— *Mr. Technology*
*Sources: [Reuters — OpenAI files for US IPO after Anthropic](https://www.reuters.com/technology/openai-files-us-ipo-after-anthropic-ai-giants-head-public-markets-2026-06-08/) (June 8, 2026); [CNBC — OpenAI confidentially files for IPO](https://www.cnbc.com/2026/06/08/openai-confidentially-files-for-ipo-prepping-wall-street-for-ai-debut.html) (June 8, 2026); [NYT — OpenAI Files Confidentially for IPO](https://www.nytimes.com/2026/06/08/technology/openai-ipo.html) (June 8, 2026); [OpenAI announcement](https://openai.com/index/openai-submits-confidential-s-1/); [Anthropic S-1 filing](https://www.anthropic.com/news/confidential-draft-s1-sec) (June 1, 2026); [TechCrunch — Google pays SpaceX $920M/month](https://techcrunch.com/2026/06/05/google-will-pay-spacex-920m-per-month-for-compute/); [NVIDIA Nemotron 3 Ultra](https://developer.nvidia.com/blog/nvidia-nemotron-3-ultra-powers-faster-more-efficient-reasoning-for-long-running-agents/); [vLLM Semantic Router v0.3 Themis](https://github.com/vllm-project/semantic-router). Private-market valuations: OpenAI $730B post-tender, Anthropic $852B post-round, SpaceX filing pending. Reuters reports a $1T target for the public offering, with a September 2026 timeline as one scenario. This post is analysis of public reporting as of June 9, 2026, 14:03 CET. Nothing here is investment advice.*