The Investor

The multiple that matters is on gross profit

Start with the arithmetic most private marks quietly skip. A company at 15x ARR on 55% gross margin is really trading at 27x gross profit, while old-fashioned per-seat software at 15x ARR on 85% margin trades at 17.6x. Same headline number, materially different asset. The empirical gross-margin band for AI-native businesses now sits at 50-60% against 80-90% for per-seat SaaS, which is a 35-40% discount to SaaS revenue multiples before anyone opens the growth-rate argument.

The comforting counter-thesis, that cheap inference restores the margin, has been tested and it failed. Token prices fell more than 95% over three years and enterprise spend on large language models still more than doubled in six months to $8.4B, because cheaper calls bought more calls. Margin expansion at this layer is an engineering program (routing, caching, small models, per-action pricing) rather than a gift from the market, and that program eats the roadmap capacity that would otherwise ship features. Salesforce, Intercom and GitHub Copilot have already moved to per-action or per-outcome pricing. That is incumbents conceding a flat seat cannot absorb usage-scaled cost.

Where the sources genuinely diverge

The Information reads Palantir as evidence that value accrues to the orchestration layer, and the capital-intensity spread is the argument: roughly 1% capex against cash flow, at a company that went from $533M to $1.9B a quarter and from 13% growth in mid-2023 to 93% now. Morning Brew supplies the companion fact nobody puts on the same slide, which is that the same equity is quoted at $123.06, down 30.8% year to date, in a July when the Nasdaq fell 3.2%. Two measurements, not a contradiction. The growth figures are the reported operating quarter, the quote is a July market mark, and nothing in the reporting revises the growth rate down. What contracted was the multiple on that growth, from a peak that assumed someone would keep paying it indefinitely.

So the honest synthesis is narrower than the bull case: orchestration is where the cash flow is, and it is not where a 2025 peak multiple survives. Both facts are on the record. Only one is in most private marks.

The growth-quality tell to screen for

Snap is the cleanest teaching case. Headline revenue grew 19% and the shares popped 10% after hours, but subscriptions rose 85% to $316M, about a fifth of revenue and roughly 58% of incremental dollars, while advertising, the actual business, grew 9%. North American daily users were flat quarter over quarter at 92 million and down 7% year over year. Any portfolio company posting subscription growth above 50% against flat-or-declining core engagement in its highest-ARPU market is running a bridge, not a re-rating.

What the smart move looks like

This may be wrong, but the edge here looks procedural rather than predictive. Private comp adjustment lags a public print by two to six weeks, and this print is the anchor every enterprise-AI deck will cite. Three ways it resolves: comps reset toward gross-profit math and the discipline pays, the multiple recovers and the discipline costs only time, or per-action pricing lifts the margin band and the denominator argument softens. Diligence that converts revenue multiples into gross-profit multiples, and demands cost-per-action disclosure instead of blended ARR, is the cheapest differentiator available this month.

Capital intensity, not model ownership, is now the sorting variable — but gross margin, not revenue, is the denominator you underwrite it on.

Caveat worth holding: one quarter is not a category. Palantir's lead rests on data-graph depth and forward-deployed people, an execution moat rather than a structural one, and the services-heavy motion is exactly what lower-touch fast followers will attack.

The moat inventory, layer by layer

Baseten's flagship technical result is real, published, and slightly awkward: roughly 20% higher throughput on GLM-5.2 than NVIDIA's own NVFP4 configurations, at matched downstream quality, verified by KL divergence against full precision. It is also reproducible, which is a lovely property in a paper and a difficult one when a decacorn multiple is attached to it. Per Latent.Space, building quantized weights, a speculator model, or a disaggregation setup each costs an experienced engineer hours to days. If nobody bothers, NVIDIA ships NVFP4 checkpoints and pre-trained speculators for free.

Beneath that headline, revenue quality bifurcates. Customers arrive on pay-per-token APIs, where providers publicly leapfrog each other from 90 to 150 tokens per second inside a day, then migrate the sticky workloads onto per-hour dedicated deployments that are far cheaper above millions of tokens an hour. The first tier is spot commodity revenue. The second carries switching costs. A comp sheet that blends them is pricing the wrong company.

Capacity, not capability, was the binding constraint

The most instructive fact of the week got a fraction of the attention Fireworks got. Moonshot released Kimi K3, a 2.8-trillion-parameter open-weight model, and simultaneously stopped accepting new consumer subscriptions because it lacked the compute to serve them. Capability parity without capacity parity converts into no revenue whatsoever. Elsewhere in the same stretch, AMD committed up to $5B to Anthropic for MI450/Helios hardware plus ROCm co-development, Meta was reported in talks to lease roughly $10B of capacity to Anthropic, and NVIDIA anchored SSI on Vera Rubin. Compute stopped being capex and became contractable rent.

The second supplier finally has a number

Wafer reported 952 tokens per second per node serving Kimi K3 on AMD MI355X, with better performance per dollar than its own Blackwell deployments, driven by memory capacity and software maturity rather than raw FLOPs. One operator, one model, one workload class, which is not a thesis rewrite. It is enough to reopen residual-value and depreciation assumptions inside any GPU-backed credit structure underwritten on single-vendor pricing power.

The move

This breaks two ways, possibly three. Serving margins stabilize because contracted capacity proves scarcer than clever kernels; or the per-token pie keeps shrinking while inference-heavy application companies bank the deflation in COGS, which is the asymmetry as it currently sits; or, the more interesting version, both happen at once at different tiers. The view is to underwrite the serving layer on contracts and fidelity, and to treat kernel-level performance claims as publishable rather than ownable. Worth noting what the alternative costs: an engineer spending days rebuilding what NVIDIA gives away free is an engineer not building the contract book. Probably wrong on timing. Less likely wrong on direction.

A 2.8-trillion-parameter model you cannot serve is not an asset — which is why contracted capacity now outranks research pedigree in model-layer diligence.

The financial position behind the exits

Alphabet is defending capital expenditure guidance of $195-205B against contractual commitments past $800B, which is a comfortable position right up until the quarter where it isn't, and this quarter it wasn't: the company posted its first negative free-cash-flow quarter since its IPO, with a securities-fraud investigation opened alongside. Sergey Brin's internal memo demanded the company urgently bridge its gap in agentic execution. A strike team was formed. The largest training run the company has announced is underway for Gemini 4. Firms that withdraw from a race do not do that.

The organizational consequence is the investable part, or rather the more interesting version of it. Research leadership wants world models and AI-for-science; senior management wants to beat rivals at enterprise agents now. Both are funded from one compute pool. Only one side controls the budget. Research VP Pushmeet Kohli put it on the record: the strategy has evolved.

A supply shock in credentialed founders

The Financial Times reports the AlphaFold team has been disbanded, with most original authors reassigned to Gemini and nearly a quarter gone from the company entirely. Nobel laureate John Jumper was moved onto a coding strike team and left for Anthropic within months, taking Jonas Adler and Alexander Pritzel with him. Noam Shazeer, a $2B acquihire and Gemini co-lead, went to OpenAI. Roughly $270B of market value evaporated across two sessions around that period.

This is probably wrong at the margins, but the thesis for a seed pipeline is that this is the cheapest frontier talent available in this cycle, and it is being absorbed by two buyers who already know exactly who to call. So the window is weeks rather than quarters, and the work is unglamorous: a named map of reassigned transfers, the quarter who already left, and adjacent AI-for-science staff, with warm introductions routed before the labs finish hiring. Hours spent building that map are hours not spent bidding on the enterprise-agent deals everyone else is crowding into.

One category the same facts make unfundable

OpenAI is running a zero-price land grab on precisely the channel Google is exiting: free access to its flagship model, Codex, its work product and 75+ life-science tools for 10,000 researchers this summer, scaling toward 100,000 through 2027, backed by more than $250M committed to external research. Zero price is not a competitor you out-execute. Any deck pitching an AI copilot for scientists without proprietary experimental data or ownership of the instrument loop is selling into a giveaway.

The discipline this imposes on infrastructure marks

There are three versions of this and only one requires action. Demand holds and the commitment book reads as foresight; demand air-pockets and a hyperscaler with four years of capex contracted and negative free cash flow finds it has less room to absorb the gap than its announcement cadence implies; or the departures stop next quarter and none of it mattered. Underwrite the middle one, because crossover and neocloud exposure downstream of that spending inherits the constraint whether or not anyone priced it in. The corollary for term sheets is narrow and cheap: a key-researcher retention clause, because this quarter showed that a single departure can be a material event at a $2T company.

The people leaving because two strategies share one compute budget are the cheapest frontier talent anyone will get to back this cycle.

Confidence note: the disbanding is press-reported rather than company-confirmed, and the market-value figure is a two-session move around a departure, not an attributable valuation of one hire.