The Investor

Where the $20B went

A run-rate is one month of revenue multiplied by twelve. On the corrected figure, OpenAI's implied monthly revenue is about $4.2B, not the ~$5.8B the $70B headline implied. That is a gap of roughly $1.7B a month, using The Information's arithmetic on the FT report. OpenAI restated nothing. The Information's Martin Peers blames "clunky calculations" by investors and reporters working from second-hand investor communications.

The valuation effect is mechanical. Take an illustrative price struck at 20x the $70B figure: it is 28x the corrected figure. Anyone bidding on OpenAI secondaries against the old number is working with a denominator that fell about 29%.

The two benchmark numbers aren't comparable

The Information reports that OpenAI and Anthropic treat the cloud partners' cut differently when they count revenue. The reporting does not say which lab reports gross revenue and which reports it net of that cut. Neither figure is audited or disclosed on a schedule. Any relative-value comparison between the two labs therefore starts with a definitional error, before a multiple is even applied.

The same gross-versus-net ambiguity runs down to the application layer. Startups that resell model capacity can book pass-through inference spend as top-line revenue.

Public infrastructure is priced on private leaks

Nvidia, CoreWeave and Nebius sold off on a correction to one unaudited number for one customer. Demand did not change; only the market's knowledge of it did. That shows how much of the listed AI-infrastructure story rests on a single lab's revenue. It also shows how exposed private neocloud and infrastructure rounds with concentrated OpenAI backlog are to the next leak, in either direction.

Peers expects IPOs to ease the problem once real numbers become visible. He also notes that some public companies have started issuing annualized estimates themselves. The run-rate convention is moving into public markets, not dying there. Companies that report audited, consistently defined revenue on a schedule should earn a disclosure premium.

OpenAI's business didn't change on Thursday; only the market's knowledge of it did, and that was enough to move listed AI infrastructure.

Governance landed the same week

The Information reports that three OpenAI safety researchers fired last week now publicly allege they were punished for raising safety concerns. Separately, Last Week in AI reports that safety lead David Robinson resigned. Writing in The Atlantic, he called OpenAI's culture broken and cited a breach of Hugging Face systems by OpenAI agents. Common Sense Media also rated ChatGPT for Teens an "unacceptable risk." None of this changes revenue. It does mean diligence on OpenAI paper now has open questions about both the numbers and the governance behind them.

The smart move

The edge this quarter belongs to investors with direct access to real financials. For marks, ask each AI portfolio company to reconcile its latest month × 12 against trailing-three-month recognized revenue × 4. The Information suggests flagging any gap above 15% for review. For new paper, write the revenue definition into the term sheet: recognized rather than annualized, net of cloud and model pass-through, with a representation on how it was calculated.

What the 90% leaves out

Three pricing mechanics sit below the sticker price.

• Tokenizer. Haiku 5.5's new tokenizer counts the same text as roughly 1.25x more tokens. That is why Anthropic's own blended estimate is ~75% cheaper, not 90%.
• Long-context step. Rates jump 5x above 100K prompt tokens, to $0.50/$2.50. Techpresso's derivation puts the effective cut above that line at only ~37.5% versus Haiku 4.5's ~$1/$5, before counting reasoning tokens.
• Reasoning. Reasoning defaults to medium and can't be fully turned off, so billed output rises with every call.

OpenAI priced the other end of the curve. Luna's higher tier starts only above 272K tokens, at $0.20/$0.75. On a 150K-token prompt, MarkTechPost's comparison (cited by Last Week in AI) shows Luna's list rates 5x cheaper than Haiku's on both input and output. Anthropic is pricing to win short requests. OpenAI is pricing to win long-context agent and document work.

Per-task math reverses the headline

When two models share a sticker price, cost per task tracks token count. Artificial Analysis figures compiled by AINews show that at max effort, Haiku uses ~162k output tokens per task against Luna's ~50k. That buys five more Index points. At equal quality (both score 38), the gap narrows to about 10%. Moving from xhigh to max effort costs ~1.8x the tokens for two more points.

The strongest real-workload data point is partner-reported. Cognition says Haiku scored 58.4% on FrontierCode 1.1 at about one-eighth of Sonnet 5's cost per task. Artificial Analysis can't yet model the tiered pricing, so independent cost-per-task figures above the 100K line are still pending.

Counting requests also hides where the money goes. Anthropic says ~90% of Haiku 4.5 requests fit under 100K tokens. Simplifying AI ran a hypothetical mix: 10% of requests at 200K tokens and the rest at 10K. In that case the long tail is ~69% of input tokens and ~92% of input spend. Companies that most need the 1M window (document review, codebase analysis, long-horizon agents) are the least likely to see the 75%.

Why the labs can afford this

Samsung guided to ~107 trillion won (~$80B) of quarterly operating profit, its fourth straight record, on AI memory demand. So the labs are cutting sticker prices while their most constrained input, HBM and DRAM, sits at record prices. Techpresso reads the tokenizer change, the 100K step and default-on reasoning as margin defense: win on the list price, recover margin through consumption.

Short-prompt apps get a real cost cut; long-context agents get a discount on the sticker and a bigger bill on the meter.

Who gains and who loses

• Gains: high-volume, short-context apps (support, classification, extraction). Also harnesses that route between models (Cognition's Devin Fusion runs Haiku under Opus 5.5), and context-management tooling that keeps workloads under the 100K line.
• Loses: moats built on inference-cost arbitrage or cheap fine-tuned small models. Also token-markup businesses, now that Anthropic's subscription API credits work in third-party harnesses.

The smart move is to diligence the distribution, not the press release. Ask for each company's request, token and spend mix above and below 100K before you credit any margin expansion.

Two tiers of AI hardware exit

Eliyan is the third AI component company in play in eight months. Together they define a strategic exit band, priced on how scarce the IP is rather than on revenue.

Accelerators clear in a different tier. The Information lists Nvidia's $20B Groq licensing deal, Etched at a $21B valuation and Cerebras at $56B through its IPO. Component buyers are paying to avoid the cost of being late, so they pay before revenue exists. Haircut Celestial's "up to" figure, because it likely includes contingent payments. The Information also warns that component founders who anchor to accelerator comps stall bank processes over bid-ask gaps.

Scarcity cuts both ways

Each closed deal removes a standalone target, which raises Eliyan's value. Each deal also removes a motivated buyer. Marvell has its optical answer and Analog Devices has its power answer. Eliyan's investor list (AMD, Cisco, Meta, Arm) is both its natural buyer list and its conflict map. Corporate investors with information rights or observer seats can chill an auction as easily as they can lead one. The process becomes a real auction only if a second strategic with an unfilled gap engages alongside the unnamed bidder.

When the insider with the best view into an asset walks at the asking price, treat that price as an ask, not a print.

Entry price is the edge

At $3B, Eliyan would be about 12x the more than $250M it has raised, which is elite capital efficiency. Flip it to entry price, though. An investor who came in at the July $1B mark needs the full ask, the price Arm declined, just to make 3x gross before preferences and dilution. A $500M post-money entry returns 3x at Empower's $1.5B and 6x at Eliyan's ask. In this layer, returns are made at entry, not at exit.

Incumbents co-opt instead of buying

The Information Briefing reports Nvidia investing in d-Matrix and chip challengers choosing partnership over confrontation. Alongside the Groq licensing deal, this points to licensing and minority stakes as a growing exit path. That matters for terms. A license-plus-team transaction may not trigger a liquidation preference unless the documents say it does.

Caveats: the reporting relies on anonymous sources, the bidder is unnamed, and neither company would comment. Eliyan's optical reliability claims are its own framing.