Deep dive: Power, Not Chips: What Now Gates AI Compute

Why today

Business Insider, citing internal Amazon emails from August, reports that AWS is taking early delivery of Nvidia GB300 racks. The program is called Early Rack Receipt. The emails projected roughly $5 billion of racks in August and about $3 billion a month after that. Some AWS sites had floor space but not enough power to switch the racks on. So finished equipment went into intermediary warehouses. One AWS team, in the emails' words, lacked "a clear forecasted path" to get incoming racks into data centers. Amazon told the publication that staging finished hardware near deployment sites is standard practice. The same week, Anthropic was named as the tenant at EverWind's planned campus in Point Tupper, Nova Scotia. That campus starts with a 95 MW phase in late 2027. Put the two stories together and you get one question. If the chips are arriving and the electricity isn't, what actually sets the pace of AI compute now?

Plainly

Start with what a GPU needs before it does any work. It needs a building, which the industry calls a shell. It needs cooling that can take the heat away. Most of all, it needs a steady supply of power at the right voltage, delivered through a substation the utility has agreed to feed. Chips run on a factory's schedule. The power side runs on a utility's schedule, and that one is slower. A utility has to study whether its transmission lines can carry the new load without putting other customers at risk. It may need new lines, new transformers or a new substation. Each one needs permits, equipment with long lead times, and often a regulator's approval of who pays. Cooling adds its own constraints. Racks this dense generally need liquid cooling, so an older hall with room to spare may still be unable to host them. Here is an analogy. Picture a restaurant group that orders a hundred commercial ovens, but the gas company hasn't connected the new locations yet. The ovens are real and paid for. They sit in a depot. What matters for revenue is the day the gas comes on, not the day the ovens arrive. Now a rough worked example, with our own assumptions labelled. Take one month's reported batch, about $3 billion of racks. Suppose the hardware is written off evenly over six years, which is our assumption, not Amazon's. Also, the newest chips lose their premium as the next generation ships. A rack stored for six months is selling into a market that has moved on by six months. So power timing affects more than the schedule. It changes the return on the chip share of the reported $220 billion capex year. Some of that money pays for buildings and power gear that come first. Some pays for chips that can only earn once the first part is done. Investors don't yet know how quickly Amazon's spending will generate revenue. That gap is the core of today's story.

What the evidence shows

Here is what is established, and how. Business Insider's report rests on internal emails and a recording of an internal meeting it reviewed. That is document-based reporting, not commentary. It shows that the early-receipt program exists, gives the dollar projections, and says some sites were short on power. The report also quotes AWS chief financial officer John Felton at a September internal meeting, telling AWS's infrastructure chief, "You've taken us free cash flow negative as a company, so congratulations." That is a joke on a recording, not an admission. But it matches Amazon's reported trailing free cash flow, which fell to negative $7.6 billion. Amazon has not denied that racks are staged. Its position is that staging is normal practice. That is a fair point. Staging some hardware before installation is routine. The open question is scale and duration, and the reporting doesn't settle either. We don't know how many racks are waiting, at how many sites, or for how long. This is also not new. In November 2025, on the BG2 podcast, Microsoft chief executive Satya Nadella said chips might sit in inventory he could not plug in. As he put it, "I don't have warm shells to plug into." Two hyperscalers describing the same constraint is meaningful evidence. On Nova Scotia, the 95 MW first phase and the 325 MW expansion come from EverWind's own announcement. They are a developer's plan. EverWind says the expansion depends on commercial, technical and regulatory conditions, and that the first phase will initially run on existing grid generation. Anthropic's lease is described only as long-term, with no term reported, but Anthropic has committed to the 95 MW first phase rather than the full 420 MW campus EverWind has proposed. Now the claim that has not been shown. One post responding to the reports argued that "NVIDIA seems to be channel stuffing," forcing Amazon and other customers to take early delivery of GPUs; channel stuffing means pushing product onto customers early to book revenue sooner. The same post goes further and blames Nvidia for higher RAM and consumer electronics prices. The posts cite no document, contract or source showing that Nvidia required early delivery. The Business Insider report, as summarized, describes Amazon's own program. As things stand, this is a hypothesis, not a finding.

How people read it

The same facts support quite different readings. The AGI-forward reading says early delivery is rational. If demand for frontier models keeps compounding, a rack in a warehouse beats a rack still waiting in Nvidia's allocation queue. Taking delivery locks in supply, and power will arrive. On this view, idle racks are the cost of being first, and scarce energized capacity will earn high prices. The AI-realist reading looks at the same warehouses and sees forecasting driven by fear of missing out. The buyer has a capex budget but no clear install schedule. Free cash flow has turned negative, and the hardware loses value as it waits. Zitron's strongest version goes further and says the supplier is managing its own quarters. That would matter a great deal if shown, and it hasn't been. The market reading is less dramatic. Power-gated supply means energized megawatts are the scarce asset. Reporting summarizing Business Insider's story says AWS raised prices for some GPU capacity blocks in early October, by approximately 15%. If that holds, scarcity is showing up in prices rather than in a glut. The policy reading focuses on who pays for the grid. EverWind has committed to fund all network upgrade costs for its project. It has also signed Canada's voluntary Responsible Data Centre Development Principles, which nothing enforces. Data centre projects keep setting provincial investment records, and BetaKit notes that Bell's $50-billion buildout broke Saskatchewan's. The policy question is whether ratepayers stay protected when a campus wants to grow fourfold in a year.

What it means for builders and investors

Our read is that power, not GPU allocation, now sets when compute comes online, at least for the largest buyers. We hold that provisionally. If AWS energizes these racks within a quarter, the story shrinks to a short timing gap. Here is what it implies for companies being built now. Inference prices may stay firmer than falling chip costs would suggest, because the scarce input is the energized rack, not the chip. Neoclouds that borrow against GPUs carry interest from delivery day, whether or not the power is on. For them, the date power arrives is a solvency question, not an operations detail. The most credible early-stage wedges shorten the time to power. Bridging and on-site generation, liquid cooling retrofits for existing halls, and load that can cut back on request all fit, as EverWind's curtailment pledge shows. Efficiency also changes meaning. When power is the limit, tokens per watt matters more than tokens per chip dollar. That favors serving-stack and alternative-silicon claims, but only if someone measures them. For a power or cooling startup, one question worth exploring might be this. Who signs for the value of a month saved, and is it written into a contract or only in a pitch deck?

What to watch

Three things would settle this. The first is Amazon's next quarterly disclosures. We'd want to see whether capex keeps rising while AWS capacity commentary still points to power. A second straight quarter of that would turn a timing gap into a pattern. The second is GPU instance pricing. If AWS and the neoclouds hold or raise prices into 2027, energized capacity is the scarce asset. If prices fall, the warehouses were simply over-ordering. The third is Point Tupper. The system studies and approvals for the 325 MW expansion are the real test, and they come well before late 2027. Any document showing Nvidia required early delivery would change our view right away.