Run two gates this quarter before you treat a hyperscaler multiple as earned. Grade duration of use, then payback on the next build.
I will not treat today’s prices as proven until both grade A or B.
I will not hide that behind a 400-row forecast.
Four markets, not one token price
Frontier labs (OpenAI, Anthropic, xAI) set the quality bar and take the highest price. Enterprise deals (Microsoft, Google, AWS plus the labs) keep the dollars: contracts, safety, lock-in. That is a compliance moat. Mid open-weight models are good enough for many jobs and pull volume down the stack. Commodity inference races to the floor. You get volume without much profit.
Token mix can slide to the cheap tiers while dollar mix stays in enterprise contracts. A one-price plan picks the wrong vendor and the wrong capacity.
Watch chip-hours and cash, not the price card
When something gets cheaper, people use more of it. That is Jevons.
The unit price of a token falls. Agents write longer answers, retry more and start new jobs.
Chip-hours, power and cash can still climb.
Falling list prices can support a build-out. They can also hide a profitless volume boom.
Two gates, four colours
Gate 1, Duration: is useful work still growing faster than two drags, models getting cheaper to run, and buildings filling up?
Gate 2, Payback: when chips are easy to buy, does the next warehouse still earn more than the cost of the money you put in?
I grade each A to D (green, yellow, orange, red). X is blue: unknown, work to do, not a fifth performance grade. If a gauge does not apply, omit it.
A on Duration: cloud is hot, jobs are thicker, use is ahead of efficiency and full buildings. D: volume is up, hours or revenue are flat, or the growth is only price cuts and in-house chips.
A on Payback: cash returns after the rush, in a loose chip market. D: the next megawatt is a value trap. You only made money because supply was tight.
What the cash is saying
Cash from 2026 to 2028 is the constraint. Value in 2030 is whether this spend comes back.
Microsoft: $332 billion sales (FY26), $116 billion capex, $67 billion free cash flow. The only giant still minting cash at peak build. $678 billion of signed work in backlog.
Alphabet: $446 billion sales (TTM), $195 to $205 billion capex planned for 2026. TTM free cash flow $53 billion. One quarter went negative. Search funds the build. Cloud has to start paying it back.
Amazon: $776 billion sales (TTM), about $220 billion capex in 2026, TTM free cash flow minus $12 billion. AWS must absorb the new megawatts. Retail cannot fund this forever.
Meta: $228 billion sales (TTM), $130 to $145 billion capex in 2026, TTM free cash flow $41 billion, second half at risk. Ads must earn a return on the new machines.
Oracle: $67 billion sales (FY26), $56 billion capex, minus $24 billion free cash flow. Signed book $638 billion. It is funding the build with debt and equity.
Palantir: about $8.2 billion sales (2026 estimate), almost no capex, $4.5 to $4.7 billion adjusted free cash flow. Software, not power plants. The price assumes years of very fast growth.
Cerebras: about $0.9 billion sales (2026 estimate), still burning cash. Not a strong stock. Payback is D. Delivery against $25 billion of signed work and 600 MW does not rescue the equity.
Those figures come from the working pack. I am not refreshing prices here.
First appraisal, not a buy list
Three gauges beat a 400-row model: is use still ahead, who owns the leftover machine, and does the next build still pay? This is an opening grade. Every arrow is still right. It is not a verdict on direction. It is not an instruction to buy or sell.
Palantir is the cleanest scorecard, not a priced buy: Use A, leftover omitted (almost no capex), next build B.
Microsoft, Alphabet and Amazon: next build C. Use is B. Leftover is B. Azure and AWS are real. Payback after scarcity is not proven. Alphabet’s cash turned negative in a quarter.
Meta: Use C, leftover B, next build C. Ads are strong. Return on $130 to $145 billion of 2026 kit is the open item.
Oracle: Use B, leftover C, next build X. Huge signed book, mix of prepaid chips. Cash and funding are the issue.
Cerebras: Use C, leftover C, next build D. Weak name, not a work in progress. Gate 2 fails. The multiple is not earned.
What to do in 90 days
Days 1 to 30: stand up the four-tier demand view and the three gauges per name. Write A to D in plain language.
Days 31 to 60: load the latest earnings, capex guides and signed-work figures. Run two shocks: open models catch up, and custom chips take share.
Days 61 to 90: one page, A to D, on whether current prices look earned, and which name is C or D on which gate. Cerebras is already D on next build.
Say yes to the working model and the two gates. Say no to treating 2026 to 2028 capex as proven profit.
The board will talk about AI as one market and one token price. It is not. Test duration and payback before you trust the price.
Request a private briefing
The 90-Day CEO Reset starts with the grade, not the pack.