Read the AMD press release closely and you find the sentence the celebration skips over: "AMD has committed to make a strategic equity investment of up to $5 billion in Anthropic in the future."
That is AMD putting up to $5 billion of its own money into the customer it just "won."
Anthropic will deploy up to 2 gigawatts of AMD's MI450-series GPUs in Helios racks. The first gigawatt starts arriving in the first half of 2027. The dollar figure attached to the servers is "tens of billions." The dollar figure flowing the other way, from AMD to Anthropic, is up to five.
The dollar figure attached to the servers is tens of billions. The dollar figure flowing the other way, from AMD to Anthropic, is up to five.
Hold that in your head while we walk through the week the bulls will remember.
Last week, the thread was hyperscalers turning on their suppliers—the idea that Microsoft, Meta, and the rest want a cheaper AI stack and will bless anyone who can build one. This week, that thread got its trophies.
Anthropic signed for tens of billions in AMD servers, per the WSJ. Satya Nadella personally blessed Helios on Azure, reportedly calling it a way to give customers "the performance, scale and choice they need." Cerebras tied its wafer-scale chips into Helios and reportedly popped about 5%. And AMD now reportedly carries a roughly $880 billion market cap.
If you own AMD as your Nvidia-challenger bet, this was the week the story came true. Eight of the world's top 10 AI companies now reportedly run workloads on Instinct GPUs. Data-center revenue reportedly hit $5.78 billion, up 57%.
Now the fine print.
Start with Helios economics. Futurum Group reportedly estimates Helios costs $5 million to $5.5 million per rack. It reportedly pegs Nvidia's second-generation Vera Rubin at $3.5 million to $4 million.
AMD's own data-center chief, Forrest Norrod, insists the pitch is "the best total cost of ownership and the lowest cost per token." Maybe on tokens per watt it wins. But the sticker is 30-40% higher than the incumbent it's supposed to undercut, and Nvidia still reportedly owns more than 95% of the market to AMD's ~4.5%.
So why are customers signing? Futurum's own Daniel Newman asked the question out loud: "Is AMD winning because of technology leadership, or simply because compute supply is so tight that anyone who can build chips finds buyers?"
That question is the whole ballgame, and the $5 billion Anthropic check is the tell.
When a vendor has to invest equity into the customer to close the deal, the deal is not purely demand pulling product out the door. It is partly dollars pushing product out. The revenue loops back through a stake in the buyer.
None of that makes the Anthropic capacity fake. Anthropic genuinely needs compute; Tom Brown said running "across a diversified range of hardware lets us map the right workloads to the right hardware." Diversification is real. But diversification is a scarcity strategy—you buy the second-best chip because you can't get enough of the first.
And the deliveries don't even start until the first half of 2027. Meta's committed six gigawatts "over time." Nadella's Azure deal ships in the second half of 2026 with no disclosed size and no disclosed price. These are options on future capacity, not shipped revenue.
The market is pricing them as certainty anyway. AMD reportedly trades at roughly 177 times earnings with a 2.47 beta. That is not a chip company multiple; it is a bet-on-AI-share-gains multiple, and you are paying full price today for wins that clear in 2027 and are partly funded by AMD itself.
Now flip to Intel, which reported the same week and looks like the photographic negative of all this.
Revenue reportedly came in at $16.1 billion against a $14.42 billion estimate. Adjusted EPS was reportedly 42 cents versus 21 expected—a clean double. Revenue grew 25%, the fastest for any quarter since 2011.
The data-center business reportedly grew 59% to $6.3 billion. Gross margin reportedly recovered to 42% from 2.5% a year ago.
Here is the part that matters for the scarcity thesis. Nobody paid Intel to take its CPUs.
There is no $5 billion equity check into a customer. No vendor financing, no revenue routed back through a stake in the buyer. There is CFO David Zinsner saying the company is supply constrained, with data-center customers "demanding more than it can produce."
That is demand pulling chips, not dollars pushing them. Intel signed 10 long-term CPU agreements, some with pricing locked in, others locking volume—customers committing because they're worried they won't get enough, not because Intel sweetened the terms with its own balance sheet.
Zinsner's line on the call was almost boring by comparison: "Customers continue to signal a strong and sustainable spending environment." No optionality, no 2027 delivery dates. Just chips selling faster than Intel can make them, at recovering margins.
And that is exactly why Intel is, on this specific question, the cleaner alternative-to-Nvidia story. The AI inference shift—models responding to requests, agents completing multi-step tasks—is a CPU-hungry workload, and Intel's Xeon franchise is catching that demand without buying it.
But do not mistake this for a verdict that Intel is winning either. The foundry is still the wound.
Intel Foundry reportedly did $5.8 billion in sales, up 31%. It reportedly landed Fortinet as a named customer—on an older process, for security chips. It did not reveal the marquee external customer investors have been waiting for.
Zinsner is guiding capex "meaningfully" higher next year, mostly for factory tooling, on the bet that TSMC is at full capacity and Intel can add cleanroom space faster than anyone. That bet is real. It is also unproven, and it is why the stock reportedly fell 28% in July even before this beat.
So here is the honest synthesis, and it cuts against both bull cases.
AMD's design wins are being priced as optionality you pay for—177x earnings, $880 billion cap, 2027 deliveries, and a $5 billion check written into the very customer it "won." Intel's wins are demand-constrained and cash-generative, real revenue at recovering margins that nobody had to be bribed to buy.
Those are genuinely different qualities of win, and on the narrow question of what's actually pulling product, Intel's is cleaner.
But both stories share one engine, and it isn't CUDA-cracking. AMD's ROCm still hasn't breached Nvidia's software moat—Counterpoint's Neil Shah says the hardware is "comparable" but "the key lies in software and optimization." Intel's Xeon strength is riding an inference boom, not a foundry breakthrough.
The engine is scarcity. Compute is so tight that a 30-40% pricier rack finds buyers, and a CPU maker with an unfinished foundry can't keep up with orders.
Newman asked whether AMD is winning on technology or on tightness. The uncomfortable answer is that the tightness is doing the heavy lifting for both companies.
Scarcity manufactures design wins. It lets vendors lock in pricing, extract equity terms, and sign 2027 commitments at today's premiums. It is a wonderful thing to sell into.
It is also the one variable in this entire story that ends. When it does, the question stops being who can build chips and becomes who built the better one—and that is precisely the question this euphoric week left unanswered.