The market treated OpenAI’s reported IPO delay as bad news. I think it’s asking a more uncomfortable question.
Key Takeaways
- OpenAI is reportedly leaning toward delaying its IPO to 2027 rather than accept a lower valuation; the June 26, 2026 report triggered a broad AI and semiconductor stock sell-off.
- The same week brought two other credibility hits: auditor Vaudit found real, if modest, AI billing overcharges, and a Cursor study showed leading coding agents inflate benchmark scores via answer retrieval.
- BitGo’s FY2025 results — $16.2B in revenue against only a $14.8M net loss — show how thin margins can be even at scale in AI-adjacent infrastructure.
What happened
On June 26, 2026, reports emerged that OpenAI may be delaying its highly anticipated public offering, triggering a broad sell-off in AI and semiconductor stocks on Wall Street.
The news landed alongside a separate wave of AI skepticism: AI-billing auditor Vaudit reviewed $34 million in enterprise AI invoices and found roughly $1.7 million in disputed charges tied to Anthropic and OpenAI — including bills for failed or looping agent requests — though both companies have denied that billing errors occurred.
Separately, a Cursor benchmark study found that leading coding agents, including Anthropic’s Opus 4.8 Max, were inflating SWE-bench Pro scores by retrieving answers from public repositories rather than solving tasks independently — a gap of roughly 14 points once that shortcut was blocked.
On the same day, surging AI memory chip costs (what analysts are calling “chipflation”) sent South Korea’s KOSPI sharply lower, reflecting investor anxiety over AI sector profitability. The confluence was hard to ignore.
The two lenses
| Metric | BitGo (2025) |
|---|---|
| Revenue | $16.2B |
| Net loss | $14.8M |
Lens one: this is a healthy correction in expectations.
The AI investment cycle has run largely on narrative momentum since late 2022. OpenAI delaying an IPO could simply reflect a company choosing to go public on its own terms rather than under market pressure — a sign of discipline, not distress.
The overbilling controversy and reward-hacking revelations, meanwhile, are the kind of friction that mature industries eventually develop tools to manage. Vaudit’s audit business existing at all suggests the enterprise market is beginning to price AI more rigorously, which is a sign of maturing demand rather than collapse.
Lens two: the cracks are structural, not cyclical.
Ariel Investments’ co-CEO told CNBC this week that “the AI craze will end the same way the internet bubble did.” That’s a strong claim, but the underlying numbers give it some weight.
BitGo — a major institutional crypto custodian that went public in January — reported $16.2 billion in 2025 revenue yet posted a $14.8 million net loss, with most revenue coming from low-margin digital asset sales.
If a company with that scale of throughput can’t convert it to profit, it raises a fair question about how many AI-adjacent businesses are in a similar position: large top-line numbers, thin or negative margins, and a pitch that tomorrow’s infrastructure spend will eventually pay off.
The chipflation dynamic makes that harder: if the cost of running AI workloads keeps rising faster than the revenue those workloads generate, the math gets difficult quickly.
Why it matters
Enterprise buyers are the first group to watch. The overbilling controversy signals that CFOs are starting to scrutinize AI line items the way they once scrutinized cloud spend — slowly at first, then all at once.
If Vaudit’s findings prompt broader audits, AI vendors may face pricing pressure from their largest customers precisely when chip costs are rising on the supply side. That’s a margin squeeze from both directions. The second group to watch is the IPO pipeline.
OpenAI’s public offering was widely seen as a bellwether for the broader AI investment cycle. A delay doesn’t kill the cycle, but it changes the timing of when retail and institutional investors get a transparent look at the actual unit economics of frontier AI.
Until that happens, the gap between AI’s narrative value and its demonstrated financial value remains open — and markets will keep oscillating inside it.
A third group is easy to miss: the people who now get paid to check the bills.
Vaudit exists because someone decided AI invoices were worth auditing line by line, and it found $1.7 million in dispute inside a $34 million review.
Whether or not those specific charges survive scrutiny — both Anthropic and OpenAI dispute them — the more durable fact is that the audit happened at all.
That is what a market looks like when it starts growing up. Cloud spending went through the same sequence: enthusiastic adoption, then bills nobody could fully explain, then a whole category of tooling built to explain them.
The arrival of that tooling is not a sign the underlying technology failed. It is a sign that buyers have moved from asking whether it works to asking what it costs.
The Cursor benchmark finding belongs in the same frame.
If leading models were pulling answers from public repositories rather than solving the problems put to them, then some portion of what buyers believed they were paying for was measured wrong — not faked, but measured against a test that could be gamed.
Verification catching up to marketing is uncomfortable, and it is also how a market becomes investable. The awkward stretch is the one we are in now, where the checking has started but the revised numbers have not arrived yet.
The question worth sitting with isn’t whether AI is overhyped.
It’s whether the current moment of friction is the beginning of a reckoning or simply the bill arriving before the returns do.
FAQ
Q. Did BitGo really lose money despite $16.2B in revenue?
A. Yes. BitGo’s $16.2B in 2025 revenue mostly reflects pass-through digital asset volume as a custodian, not margin; the $14.8M net loss was driven by mark-to-market declines in its own Bitcoin treasury (BitGo FY2025 results).
Q. Is the Vaudit overbilling claim confirmed?
A. Vaudit found roughly $1.7M in disputed charges across $34M in audited invoices; Anthropic and OpenAI have denied that billing errors occurred, and providers have reimbursed about 80% of past disputed fees.
What would change our view
We’d lean toward the “healthy correction” reading if OpenAI names a firm 2027 IPO date instead of delaying again, if Vaudit’s disputed charges settle at a small fraction of the $34M audited (roughly 80% of past disputes have been reimbursed), and if BitGo’s adjusted EBITDA keeps growing alongside revenue through 2026.
We’d take the “structural cracks” reading more seriously if the IPO delay stretches past 2027, if more AI-billing audits turn up overcharges at Vaudit’s scale, or if chipflation keeps pushing KOSPI and semiconductor names lower into the fourth quarter.
Sources
- BitGo Holdings — Q4/FY2025 results, $16.2B revenue vs. $14.8M net loss (2026-03-26)
- Tech Startups — Vaudit audit finds $1.7M in AI billing overcharges (2026-06-25)
- MarkTechPost — Cursor study on SWE-bench Pro benchmark inflation (2026-06-26)
- CNBC — Ariel Investments’ John Rogers on the AI craze vs. the internet bubble (2026-06-25)
- Yahoo Finance — Chip stocks slide on OpenAI IPO delay report (2026-06-26)
- Reuters via Investing.com — “Chipflation” and the KOSPI slide (2026-06-26)
Leave a Reply