The Investor

The bull case is in the compute line

Most coverage skipped compute intensity. Augment's analysis of the Reuters-reviewed figures puts 2025 compute and infrastructure spend at $7.33B on roughly $4.6B of revenue. In 2024 it was about $2.44B on ~$383M. Compute per revenue dollar therefore fell from about $6.37 to about $1.59, a roughly 75% drop, while revenue grew about 12x and compute grew 3x.

There is a caveat. Each new revenue dollar still needed about $1.16 of new compute, so the business does not yet fund itself at the margin. The compute figure also mixes training spend for future models with inference spend for current customers, so it is not gross margin. How the public S-1 splits those two is the disclosure that will show whether the ratio keeps falling.

Sources disagree on the revenue base, and that is the valuation question

Augment puts the $2T+ target at about 435x trailing revenue, against about 210x at the $965B May round. It argues $2T only works if the S-1 shows a run-rate of roughly $20–40B. AINews reports about $11.5B of Q2 2026 revenue and ARR above $65B, which would make $2T about 31x ARR. Q2 annualizes to about $46B, roughly 40% below that ARR claim. The gap means either sharp growth after Q2 or a generous definition of ARR. Unwind AI says no revenue figures leaked at all. Treat any 2026 run-rate as unverified until the public filing.

The $518B looks more rigid than it is, and that shifts risk to suppliers

The Information's reading of the prospectus changes the risk map. Anthropic has committed up to $84.5B to SpaceX through 2029 for Nvidia-based compute, and most of it can be cancelled on 90 days' notice. For Anthropic, a headline fixed obligation works like a variable-cost option. For SpaceX, and for any supplier on similar terms, that backlog is closer to a rolling purchase order. Augment separately calls the $518B its lowest-confidence figure and reads it as multi-year contracted commitments, not one year of cash.

Anthropic has protected itself against a delayed listing; a supplier booking lab backlog at face value is holding the option Anthropic kept.

The reporting doesn't say whether SpaceX's termination rights are mutual. That matters because SpaceX now owns Cursor and runs Grok Bot, so it competes with Anthropic.

Three more disclosures to price

• Concentration: about $1.1B, nearly a quarter of 2025 revenue, came from two unnamed customers. The Information AM notes that their identity decides whether this is sticky revenue through hyperscaler channels or spend from app-layer startups that can move.
• Cadence: the risk factors say the business requires a "continuous and overlapping cadence" of model releases. The filing surfaced the same week OpenAI shelved GPT-6.1 Astra after safety testing. A $2T price quietly assumes releases never slip.
• Control: Anthropic is reportedly seeking Palantir-style voting control for its seven co-founders. That is more likely to affect price than to kill deals.

The smart move

Augment expects the listing after the November midterms. The confidential filing has to go public at least 15 days before the roadshow, so the real numbers are weeks away. Until then, the only defensible mark is the priced round. Also flag portfolio companies that carry SAFEs, convertible notes or warrants classified as liabilities. Their Q3 markups will book the same kind of non-cash GAAP loss that inflated Anthropic's headline.

Oura failed on structure, not macro

Oura blamed 'market uncertainty'. The Information Dealmaker's numbers suggest a less flattering culprit, or rather a more specific one: Oura's own choices. It asked for about 9x estimated 2026 revenue, roughly Apple's multiple, for a business that made $60.8M of net profit on $1.2B of revenue over the nine months to June (about a 5% margin) and lost money in the June quarter. Forerunner planned to sell its entire ~9% stake in the offering, more than 28M shares worth about $1.2B at the midpoint. Augment counts 73% of the 50M shares on offer as secondary.

The deals that cleared looked different. No major holder sold in Cerebras' $56B IPO, and Valor and DFJ Growth kept every SpaceX share in June. Insider selling is sometimes just liquidity, but IPO buyers now price it as a primary input, and a founder or lead investor cashing out at listing is treated as a bearish signal.

The AI build-out hits the same filter

SB Energy, SoftBank's data-center power developer, filed publicly on September 1 and still hasn't started marketing. It has no operating data centers. Nscale is pitching $35B on data centers that aren't built. The Information reports that debt investors have become more selective about financing data centers and chips. One market tightening is a pause, and equity and credit tightening on the same asset class at once is how a pause becomes a repricing. The counter-case is that private buyers still pay for leadership. OpenAI's $30B raise at up to $1.4T works out to about 20x its ~$70B annualized revenue.

The credit tape says the tenant layer is the weak link

Tech now dominates convertible issuance, per The Information's capital-markets reporting:

• Tech is about 60% of 2026 US convertible issuance, $78B through September 11, up from about 44% last year, and that excludes CoreWeave's $4.2B sale.
• AI-specific debt costs are rising, and Jane Street-linked data-center debt is souring.
• The GPU futures effort has stalled at the CFTC. Lenders still have no market hedge on what used GPUs will be worth.

Some neoclouds stack three layers of fixed obligations on the same short-dated compute revenue: GPU debt through special-purpose vehicles, project finance for power, and convertibles on the corporate balance sheet. When the least-bad instrument in that stack gets crowded, the next stop is dilutive equity. Morning Brew has the 10-year Treasury at 5.24%. Long-duration growth marks reprice off that number before anything else does.

Where the premium goes

This is probably wrong at the margin, but the premium appears to be clearing where capital avoids the public-story test altogether. In the Williams power deal, the disclosed figures imply about $10.8B of value on about $9.0B of project cost. That roughly 1.2x markup accrues to the developer's equity, not to the project investors. IDF argues its power blocks pay back in 6–8 years. That holds only if contracted compute revenue lasts that long. Behind-the-meter assets exist because grid interconnection takes years, so the option to sell power back to the grid deserves a steep discount.

Until data-center debt loosens, only energized capacity and clean insider behavior clear the public market. Everything else gets priced by whoever is still willing to lend.

The hyperscaler playbook moves up to the model layer

AWS and Azure won their marketplaces by letting customers spend commitments they had already made, which is the dullest and most effective trick in enterprise software. Both frontier labs now run it. AINews reports that OpenAI's B2B Marketplace lets enterprises apply OpenAI commits to open models served through Baseten, so owning the prepaid budget pays even when the workload runs on someone else's model. For the buyer, anything outside the commit competes with a line item already paid for.

Sign in with ChatGPT solves a margin problem and creates a control problem

Inference has long been the gross-margin drag for coding and agent tools. Sign in with ChatGPT moves that cost onto the user's ChatGPT plan, and OpenAI takes the billing relationship in exchange (usage caps and pricing power come attached). It has already shown how fast it uses them. Consumer plans were re-tiered to usage multipliers: Plus 1x, Pro 100 5x, Pro 200 10x and a new Pro 500 at 25x. AINews estimates the change roughly halved the old Pro 200's value, and Techpresso reports the reopened $200 Pro plan came with API credits halved. One tier change can reset a portfolio company's unit economics overnight.

Cheaper models feed the rail

Deflation makes the bundle cheaper to subsidize. OpenAI's GPT-6.1 Sol costs one-fifth of Astra, at $2/$10 per million tokens. In an independent planted-bug test reported by AINews, Sol found 44 bugs for $6.56 while Claude Opus 5.5 found 41.7 for $58.53, or about $0.15 versus $1.40 per bug. Sol also emits 10–30% more output tokens, and results varied between test setups. Anthropic is holding Sonnet 5.5 at $2/$10 while claiming up to 30% lower cost per task. App-layer margins widen, but only where pricing does not hand the savings straight to customers.

Seats are being repriced as capacity

Turing Post notes that ChatGPT in Slack and Teams no longer needs a license per participant, and Dots scales delegated work with plan tier rather than seats. Casey Newton's hands-on test estimated about two hours of back-office work for about 15 minutes of his attention. That is one user's field report, and the ~$100/month price was his guess. Still, seat-priced SaaS sitting next to agent workflows faces pressure on its valuation multiple, not merely its feature list.

Where the rail is weakest

Neither lab has disclosed take rates, drawdown caps or eligibility rules, and those terms decide whether the marketplace is a channel for portfolio companies or a tax on them. This may prove wrong at the edges, but the squeeze looks horizontal: meeting notes, docs, code review, generic browser agents. Systems of record, regulated workflows and vertical data depth still sit outside the bundle. AINews also notes the new Decisions API ships without calibration, which leaves an opening for calibrated decision models.

Horizontal AI apps are now competing with prepaid lab budgets; vertical depth and system-of-record status are what keep a product off the bundle's menu.