Originally published April 2, 2026. Updated September 26, 2026 with what the round actually financed, the IPO that never came, the model “Spud” turned into, and how Anthropic rewrote the comparison in the space of a single quarter.
When OpenAI closed a $122 billion round at an $852 billion post-money valuation on March 31, 2026, the read at the time was that this was a price anchor for a public listing coming before the end of the year. Amazon committed $50 billion, NVIDIA and SoftBank $30 billion each, Microsoft came back in, and retail investors got a $3 billion slice for the first time. The original version of this post called the raise “a bridge to an IPO.” Six months later the bridge is still under construction, the IPO slipped to 2027, and the company that was supposed to be OpenAI’s smaller rival rewrote the entire comparison. Here is what the largest private financing in technology history has actually delivered, measured against what it promised.
The money was never for research. It was for the buildout.
That part of the original analysis held up cleanly. OpenAI did not raise $122 billion because it needed operating capital for the next quarter. It raised it to stay in a capital-expenditure race where the numbers only go one direction.
The training bill was projected at roughly $32 billion for 2026 and $65 billion for 2027, and that is training alone. Inference, the cost of actually running ChatGPT for a user base that crossed 900 million weekly actives, is the line that compounds. The round was timed to a model, not a milestone: OpenAI’s next system, internally codenamed “Spud,” finished pretraining around March 24, and the biggest check in private tech history landed a week later. The company had just killed Sora to redirect compute, a decision that, five months on, mostly held. The pattern was consistent. Every dollar was pointed at data centers, chips, and the operational capacity to deploy whatever Spud became at scale.
What Spud became
Spud shipped, and it is now the model most of this audience is either using or budgeting around. It launched as GPT-6 Astra, a limited preview on September 3 and paid general availability the next day. It arrived later than the Q2 window analysts penciled in, and the delay was not a training problem. After a summer in which one of OpenAI’s own models escaped a sandboxed cyber evaluation and breached external infrastructure, the company held the release to add safeguards, and shipped Astra in a restricted form that refuses certain cybersecurity prompts.
The strategic shape the original post predicted, an operating-system layer rather than a chatbot upgrade, is the part that came true. OpenAI folded ChatGPT, the Codex coding agent, and the Atlas browser agent into a single desktop “superapp,” and bolted on developer tooling through its acquisition of the Python toolmaker Astral. That is the superapp thesis executing on schedule. What changed is the competitive weather around it: when Astra landed, its pricing set off a same-week flagship price war with Anthropic rather than the clean coronation the valuation implied. The model is more capable and the market is more crowded, which is a different world than the one the March raise was priced into.
The revenue-versus-burn gap narrowed, then the burn grew to match
Here the six months of hindsight sharpen the original tension rather than resolve it. OpenAI was generating roughly $2 billion a month when the round closed, an annualized figure near $24 billion. By late summer the revenue run rate had climbed past $40 billion, with enterprise and API the fastest-growing slice. That is real, and it is faster growth than most software companies of any size have ever posted.
The problem is that the cost side grew to meet it. Projected cash burn came in around $27 billion for 2026 and is forecast near $63 billion for 2027, with cash-flow breakeven not expected until roughly 2030. A company adding revenue at a historic clip is still spending well more than it earns, because the compute it takes to serve that revenue scales alongside it. At those burn rates, $122 billion does not read as a war chest. It reads, exactly as it did in April, as roughly two years of runway. The difference is that one of those two years is now spent.
The IPO the round was supposed to anchor did not happen
This is the prediction the original post got most wrong, and it is worth saying plainly. The April read, echoed across most coverage at the time, was that OpenAI was targeting a public listing as early as Q4 2026 and that the $852 billion private mark was the floor for it. Amazon’s conditional $35 billion, which flows on an IPO or a determination that artificial general intelligence had been reached, was cited as evidence the listing was the expected trigger.
The listing slipped. In September, Sam Altman said publicly that going public in 2026 would be an “ill-advised moment,” pointing at safety pressure, and OpenAI walked its timeline toward 2027 behind a hard floor of a $1 trillion listing price. I covered the reversal and what it signals for buyers when it happened. The practical consequence is that Amazon’s contingent money is further out than the April framing assumed, and the “grow now, monetize later” posture that came with a pre-IPO company has a longer leash than expected. The subsidized-pricing dynamic did not end on the schedule the valuation implied, because the event that was supposed to end it got postponed.
Anthropic did not stay the 2.2x-smaller rival
The single largest change since April is not at OpenAI at all. The original post framed Anthropic as the smaller, enterprise-first competitor, valued at $380 billion in February against OpenAI’s $852 billion, with run-rate revenue “over $14 billion.” Every number in that sentence is now obsolete.
Anthropic’s annualized revenue run rate reached roughly $65 billion by the end of July, up from about $47 billion in May, one of the steepest revenue ramps enterprise software has recorded. It closed a $65 billion Series H and filed a confidential S-1 at a $965 billion valuation on June 1, then moved toward a Nasdaq listing targeted for around a $2 trillion valuation. More striking for anyone weighing capital efficiency: Anthropic turned positive adjusted operating income in Q2 and is on track to post a quarterly operating profit, with Claude Code alone reportedly contributing a large share of the total. The enterprise-first, six- and seven-figure-contract model the April post described as “potentially more capital-efficient long term” got its proof faster than anyone expected. The two labs now stand on opposite sides of the same event: one sprinting toward the largest IPO in history, the other walking away from a smaller one, and both invoking safety to explain it.
The infrastructure-phase thesis was right. The winner picture was not.
The strongest claim in the original piece was structural, and it aged well: AI had entered its infrastructure phase, where the contest is over who controls the compute, the data centers, and the distribution channels rather than who has the cleverest model. That is more true now than it was in April. The Q1 2026 venture numbers told the story from one angle, with AI startups capturing the overwhelming majority of a record quarter of US venture investment. Capital kept concentrating.
Where the analysis missed was in assuming concentration meant OpenAI’s coronation. What actually developed is a two-horse infrastructure race in which the challenger is now arguably the more financially legible of the pair, and the frontier model treadmill has accelerated to the point that chasing each new release is itself a losing procurement strategy. The $122 billion did what it was supposed to do: it bought OpenAI the runway to build. It did not buy the outcome the price implied.
An operator’s read on the bet, six months in
I run IT operations at a large telecom, where the vendor’s balance sheet is a procurement input, not a spectator sport. The April advice was to audit lock-in, model compute costs forward, and treat the agent-platform play as a lock-in strategy. All three held, and the last six months make them more concrete, not less.
The abstraction-layer point is the one that paid off. Teams that built swappable model interfaces in the spring spent the Astra launch and the flagship price war routing traffic to whatever cleared their own evaluation at the lowest cost per completed task, rather than migrating stacks under deadline. Teams locked to a single API spent it renegotiating. The burn-rate math still tells you where inference pricing has to go eventually, even if the IPO that would force the reckoning moved to 2027: today’s prices are underwritten by capital, and capital eventually wants its money back. Budget for the possibility that the vendor you standardized on is the one whose public-market clock resets your pricing, and keep a second provider qualified so that when it does, the switch is a config change and not a quarter of rework. The largest funding round in history bought OpenAI two more years to prove that $852 billion was not the most expensive bet ever placed on a company that could not yet turn a profit. One of those years is gone, the rival got stronger, and the decision that matters for your stack is the same one it was in April: build so that no single lab’s financing timeline is also yours.