Abundance still has a power bill

A down week in the AI complex is supposed to be a verdict. The week of 17 to 21 August was a tape. Technology, the XLK fund, lost 3.5%. The Nasdaq 100 lost 2.45%. Friday's own session closed +0.11% and +0.33%. No margin printed. No capital-spending guide printed. The series that matter are quarterly. What did print, in July, is the size of the bill.

Amazon spent $54.2 billion on cash purchases and additions to property and equipment. Alphabet spent $44.9 billion. Microsoft spent $35.8 billion. Meta spent $30.1 billion. The four investor-relations and 10-Q cash-flow lines sum to $165 billion. Platformonomics prints $172 billion. Axis Intelligence prints $170.1 billion, equal to 99% of the four companies' operating cash flow in the same three months. Those third-party cuts mix leases and different bases. Combined 2026 guidance after the July prints sits around $735 to $750 billion.

A cheaper token is not a cheaper rack

Flagship API list prices, three official pages fetched 22 August 2026. OpenAI gpt-5.6-sol at $4 input and $20 output per million tokens on the standard Flagship table. The $5 and $30 pair is gpt-5.5 under 272k, not Sol. Anthropic Claude Opus 4.8 at $5 and $25. Opus 5 sits on the same list. Fable 5 at $10 and $50 is a different rung. Google Gemini 3.1 Pro Preview at $2 and $12 at or below 200k context, and $4 and $18 above it. DeepSeek V4 Pro is an order of magnitude cheaper on the published sheet. Cached input and batch cut those numbers again. The frontier is a ladder, not a monopoly price.

Token list prices can fall 60 to 75% while the cluster bill rises, because memory and power do not trade on the same sheet as output tokens. A cheaper completion is not a cheaper rack. Amazon moved a $200 billion cash figure to about $220 billion and said memory did part of the work. Microsoft and Meta had already cited component costs earlier in the year.

Amazon is the exposed invoice

Andy Jassy's sentence is the one that belongs on this page. Even at $220 billion, Amazon will not have enough capacity to meet 2026 demand, and he believes the same will be true in 2027. Alphabet moved to $195 to $205 billion. Meta narrowed to $130 to $145 billion. Microsoft's calendar figure printed near $175 billion only because more data-centre leases are being classified as operating rather than finance leases. The company said the underlying build is unchanged, and the next quarter's capex, including that shift, would exceed $50 billion.

Amazon's AWS backlog of $496 billion, up from about $364 billion a quarter earlier, is the one number on the build that looks like an invoice. Jassy said Trainium2 is essentially sold out and much of Trainium4 is already reserved. Alphabet cannot point to an equivalent contracted figure. Meta cannot point to one at all, which is why the call spent time on inbound interest in selling compute externally. A backlog is not a gross margin. It is evidence that at least one of the four is not building into a void.

Cash flow is now the constraint

Axis Intelligence's 99% reading of capital spending against operating cash flow for the four names in the second quarter is the sentence the equity market has been refusing to write. Alphabet and Amazon each spent more on property and equipment last quarter than they generated in operating cash flow, on the cuts that have been circulating since the July calls. Meta's free cash flow after finance-lease principal was $784 million. Microsoft is the exception: operating cash flow still covered cash purchases, with free cash flow near $19.6 billion. A build that cannot be funded out of operations is a build that will be funded with debt. That does not end scarcity. It changes who gets paid if the cost of capital rises: the lessor, the bondholder, the memory vendor, before the application vendor does.

Three layers, three answers

Picks and shovels, accelerators, memory, networking, shells, power, have captured nearly all the equity value so far. Their pricing power is a queue. High-bandwidth memory and turbine slots are allocated, not shopped. When unit volumes double and vendor gross margins hold, scarcity has not ended. When volumes double and margins fade, the supply response arrived. A week in XLK does not settle that.

The model layer is the one the market watches. Benchmarks are the tell it watches. They are the wrong tell. When the frontier is a commodity, the margin goes to whoever has distribution, not whoever has the best benchmark.

The application layer is where abundance either shows up as the vendor's margin or as the customer's saving. ICONIQ's 2026 State of AI snapshot put inference at roughly 23 cents of every AI-product revenue dollar at scaling-stage business customers, and average AI-product gross margin at 52%, better than 41% in 2024 and 45% in 2025, and still a floor that would have been a crisis print for a 2019 software underwriting model. Snowflake's product GAAP gross margin is 71%, non-GAAP 75%, down from a world that talked about 78%. ServiceNow's fiscal 2026 GAAP subscription margin guide is 75%, and the company has named AI and hyperscaler partnerships in the same sentence.

Private equity's old software move was to buy the 80% margin and cut the costs that sit above it. Inference sits below it. You cannot fire a token. Caching and distillation are real. They do not restore the underwriting law. The dollars rotate from engineers and support into cost of goods. A roll-up that treats that as a temporary tax will cut the product when it cuts the bill.

Hosting cost as a share of revenue is the next quarter's tell. Several first-quarter 2026 management discussions already isolate 4 to 9% of revenue as inference. The names that have not started that sentence are the names whose 80% margin is a timing difference. Auditors will force the classification: inference required to deliver the subscribed feature is cost of goods, not research.

What a Friday cannot do

XLK down 3.5% with the Nasdaq is the short truth. It will be quoted as a refutation of a compounding build. It is not. A refutation would look like this: the four companies' cash purchases go flat to down year on year for two quarters with no corresponding rise in disclosed AI revenue; or flagship list prices stop falling for four quarters; or vendor gross margins hold while unit volumes double. A Friday session is none of those.

NVIDIA on a Wednesday night is an input to whether the build is still sold out. Earnings per share and the data-centre number are not a software gross-margin event. Do not write a sentence that uses one to explain the other.

Power is the one layer where scarcity is not ending on a software clock. That argument lives on the power page. A campus that cannot point to a turbine is not an application-margin event.

IT-services revenue per employee is the clean public series for labour substitution. Rising with flat headcount is the substitution arriving. Services firms can look healthy on revenue while the headcount line is the whole story. Official productivity statistics are last to know. You see the computers everywhere except in the productivity numbers. That line is a decade-scale series. It will not settle this page inside a quarter.

The honest uncertainty is who captures the abundance. If application gross margins expand, abundance is real and shareholders capture it. If they compress, abundance is real and customers capture it. That is a fine outcome for the world and a bad one for the shares. Amazon's backlog is a fact about one invoice book. Snowflake's 71% is a fact about one product margin. Those two sentences can live in the same year.