Amazon's $25 Billion Chip Run Rate Is Proof the Inference Shift Has Arrived

Amazon's $25 Billion Chip Run Rate Is Proof the Inference Shift Has Arrived

The market read Amazon's second-quarter report the way it reads any big number: custom silicon crossed a $25 billion annual revenue run rate, growing at triple-digit percentages year over year, and the stock jumped roughly 13% in premarket trading. That framing converts Amazon into "now a chip company," which is the wrong lesson. The right lesson is that the inference transition — the shift from training AI models toward running them, where cost per token and watts per output decide which infrastructure wins — has arrived inside the largest cloud on earth, and Amazon built its own silicon specifically to own that cost curve.

What the $25 Billion Actually Contains

Start with the growth context, because it is the signal I trust most. AWS grew 37% year over year in the quarter, its fastest growth in 18 quarters, accelerating to a $169 billion annualized run rate. That kind of sequential acceleration at cloud scale is not a rounding error; it means the AI buildout is flowing through to real, billable demand — and it is happening while Microsoft and Google are fighting over the same enterprise budgets.

The $25 billion figure is important, but it is a bundle. It combines Trainium (the AI training and inference chip), Graviton (the ARM-based CPU for general cloud computing), and Nitro (the networking and security silicon underneath all of AWS). The bulk of that run rate is not AI-specific — Graviton alone has been a quiet, profitable workhorse for years, and Amazon does not disclose how the $25 billion splits between the AI chips and the CPUs. That data gap matters for discipline: the story here is not "$25 billion of AI chips," it is "$25 billion of custom silicon of which the fastest-growing, triple-digit piece is Trainium," with AWS's Bedrock service running most of its inference workloads on that AI silicon.

That is the distinction that changes the judgment. If Trainium were already a $25 billion business, the inflection would be priced. Because it is a smaller base compounding at triple-digit rates, the market is still pricing Amazon as a consumer retailer with a good cloud, not as a company whose silicon sits on the right side of the next compute transition.

Why Inference Rewards Custom Silicon

The reason this is not a niche cost-saving story is the market structure underneath it. In the training phase of the AI cycle, Nvidia's CUDA software moat is formidable, and general-purpose GPUs dominate. In the inference phase — which is where economics increasingly concentrate, since every model in every product has to keep running — the decisive metrics are latency, cost per token, and watts per output. Those are exactly the dimensions where a purpose-built ASIC beats a general-purpose GPU, and they are exactly where the CUDA moat is thinnest.

The architecture data backs this up. Trainium3 delivers up to 40% better price-performance than Trainium2, and over five times more output tokens per megawatt — meaning the same power budget generates five times the inference. Power has replaced raw compute as the industry's binding constraint; hyperscalers now discuss megawatts before they discuss chips. A five-fold efficiency jump per watt is the kind of architectural generation gap that changes cost structures, not just spec sheets. When AWS goes to a customer and can undercut Nvidia-based instances on price while keeping margin, that is not a feature bullet. It is the entire competitive strategy of cloud.

The Supply Chain Is Already Voting

This is where the demand signals, which lead earnings data, matter most. Anthropic has committed to using up to five gigawatts of current and future Trainium generations and runs Claude on more than one million Trainium2 chips. OpenAI has committed to consuming two gigawatts of Trainium capacity beginning in 2027. Meta has signed up for tens of millions of Graviton cores. These are not speculative endorsements; they are contracted capacity from the two model labs that matter most and the largest frontier-model spender.

The same lens that makes me treat surging supply commitments as leverage risk cuts the other way here. When customers commit gigawatts years in advance, they are de-risking the enormous capital Amazon is pouring into data centers — someone has already agreed to pay for the capacity before it is built. That is the difference between a speculative capex bet and a contracted one.

The Tell: Trainium May Not Stay Inside AWS

Here is the part the stock is not paying for. On the earnings call, Andy Jassy said there is a "real chance" Amazon will sell Trainium outside AWS, and the discussion now extends to selling the chips directly to third-party data centers separate from the cloud. That is the market structure flip within Amazon itself: from custom silicon as an internal cost lever to custom silicon as a merchant product competing head-on with Nvidia and AMD for the inference TAM.

Chew on what that means. A hyperscaler with a 37%-growing cloud does not need to sell chips to make its cloud work. It sells them when silicon has become strategic enough to attack the merchant market directly — and the merchant market's weakest flank is exactly inference, where Nvidia's software moat is at its most exposed. If Trainium ships outside AWS, Amazon stops being a customer negotiating against Nvidia and becomes a competitor in Nvidia's most defensible economic territory. The option value alone justifies a meaningfully different multiple than the one the stock carries today.

However: Demand Is Not the Problem

The balance sheet is the part of this story that keeps me from endorsing a blind "load up." Amazon now expects 2026 capital expenditure of approximately $220 billion, raised from around $200 billion, and it raised the number partly because memory prices climbed. Management was direct that even at that pace, capacity still won't satisfy demand — the strongest possible demand signal, and simultaneously the clearest statement that the company is paying for a year of growth before it earns it.

The numbers bear this out. Trailing twelve-month free cash flow — operating cash flow minus capital spending — has flipped negative, to roughly negative $11.6 billion on about $161 billion of operating cash flow and $173 billion of capex. Essentially every dollar of operating cash flow, plus more, is being reinvested into compute. This is the supply-commitment dynamic I have flagged for years: demand is robust, but a company that is free-cash-flow-negative must keep the capital markets and the multiple on side until the capacity converts to revenue and profit. The difference here is that we can see the contracted demand on the other side of the spending.

What This Means for the Allocation

The price context matters before any sizing decision. The stock trades near $258, up about 24% over the last 120 days and roughly 12% year to date — the earnings pop was digested a few weeks ago, so the milestone is not a fresh catalyst to chase, and it is also not a crowded trade that has front-loaded the returns. On EV/EBITDA, Amazon sits near 16.6 times, below Microsoft near 18.3 times, Alphabet near 23.6 times, and Nvidia near 31 times — cheap relative to the same part of the AI trade, for the fastest-accelerating cloud of the group. A forward price-to-earnings of roughly 40 times looks steep, but I lean on growth acceleration and cash multiples for a reason: a P/E on a quarter that carried one-time gains tells you little about a company reinvesting nearly all its cash flow.

So where does this leave the decision? This strengthens the case for Amazon as a core allocation in the AI cycle — not because of a chip run rate, but because that run rate is proof AWS has the cheapest structural inference cost curve of any hyperscaler, the fastest-growing cloud, and an unpriced merchant-silicon option. The thesis would break if Trainium's growth decelerates before external sales materialize, or if capex keeps climbing without free cash flow turning positive through 2026 and into 2027 — then the leverage side wins and the allocation should shrink. On a multi-year horizon, with Trainium3 ramping and Trainium4 already in development, the inference cost advantage compounds in the back half of this decade.

The debate is not whether Amazon is on the right side of the inference transition — it plainly is. It is whether a stock carrying negative free cash flow and a $220 billion capex bill is the best use of capital compared with the alternatives in the same phase of the cycle. For investors who can hold through the capex digestion, this is a position to own and add to on weakness; the $25 billion milestone confirms the direction rather than marking an entry to chase. The load-up instinct is right in direction. Size it for the fact that, like most of the AI infrastructure trade right now, the payoff is back-half weighted.