A Qualcomm AI rack pulls 160 kilowatts, the same envelope as a high-end Nvidia rack, and that single number is what should bother you about the low-power story Cristiano Amon is selling into Europe’s AI gigafactories. He signed a letter of intent with the European Commission this week, posted the tidy paragraph about energy-efficient, open, sovereign infrastructure, and it reads like a supplier planting a flag before the bids close November 12. I broke down the seven-gigafactory call when it landed, and Qualcomm got exactly one clause in that piece, sitting next to AMD and Nvidia as an LOI signatory. This is me going back for the clause I skipped, because the Qualcomm AI gigafactories angle turns out to be the most interesting technical mismatch in the whole program.
The gigafactories exist to train frontier models, the largest LLMs, the trillion-parameter systems that need a hundred thousand accelerators grinding for weeks. Qualcomm does not make a training chip. Every piece of data-center silicon it has shipped or announced, from the old Cloud AI 100 to the new AI200 and AI250, is built for inference: running a model that already exists rather than building one from scratch. Nvidia and AMD supply the GPUs that do the training. Qualcomm supplies the chips that serve the answer afterward. For a program whose entire justification is sovereign frontier training, that is a strange thing to volunteer for, and nobody in Brussels has said which half of the building Qualcomm is meant to fill.
The road back from Centriq
Worth remembering that Qualcomm has stood on this exact spot before and walked off it. In 2017 it shipped the Centriq 2400, a 48-core Arm server CPU it had tested with Meta, pointed straight at Intel’s data-center monopoly. Inside a year it was dead, cut in the panic of Broadcom’s hostile takeover attempt, and the first data-center dream went into the drawer marked expensive mistakes. What survived was quieter: the Cloud AI 100, an inference accelerator running on the same low-power DNA Qualcomm had spent two decades sharpening in phones. The Hexagon NPU inside every AI200 today descends directly from the neural block that started life sipping milliwatts in a Snapdragon, and that lineage is the honest core of the low-power pitch.
The credibility for the current push arrived on a $1.4 billion check. Qualcomm bought Nuvia in 2021, a startup founded by Apple silicon veterans to build server CPUs, and Arm sued within months over whether the architectural license carried over. Qualcomm put the Nuvia cores, branded Oryon, into laptops first with the Snapdragon X Elite, and only after a jury sided with it on the core licensing question in late 2024, in a trial that ended in a mistrial and is still grinding through appeal, did the server ambition climb back out of the drawer. At Computex last May, Amon said it plainly from the stage: Qualcomm is expanding into the data center, with an Oryon CPU built for clusters of inference that are high performance and very low power. He bolted it to Nvidia’s NVLink Fusion the same day, making Qualcomm one of two launch partners allowed to wire custom CPUs straight onto Nvidia’s GPU fabric.
Then it moved quickly. October brought the AI200 and AI250 as real rack-scale products with real numbers: 768 GB of LPDDR per AI200 card, which sounds like a lot and is the actual point, because LLM inference hits a memory-capacity and bandwidth wall long before it runs out of compute. The AI250 pushes further with a near-memory computing architecture Qualcomm claims delivers over ten times the effective memory bandwidth at much lower power. In December it closed the Alphawave Semi acquisition for the high-speed interconnect IP it was missing, and put Alphawave’s CEO in charge of the whole data-center business. So what Amon is pitching Brussels is not a lonely accelerator card anymore. It is a stack coming together: Oryon for the CPU, Hexagon for the NPU, Alphawave for the wires, Nvidia’s fabric to knit it. That last dependency matters, and I will come back to it.
Line Amon’s words up against the silicon and the fit is better than I expected, on exactly one axis. He leads with energy-efficient and low-power, and that maps onto the single European constraint I keep hammering. The reason Europe’s gigafactory bids are fragile is not the chips, it is the electron gap: grid connections that take three to five years to deliver, on the continent carrying the highest industrial power prices of the three racing. Training is a spiky one-time burn, but inference is the load that never switches off, the round-the-clock draw that decides whether a campus’s power budget survives a decade. When your binding constraint is megawatts you can actually get connected, the efficiency of the inference tier is exactly what determines how much useful AI you extract per megawatt. Amon is aiming at the right constraint, just not the headline one, and I doubt that is an accident, since inference is the only thing Qualcomm has to sell.
Low power at Qualcomm does not mean the rack sips electricity, and the 160 kilowatts is the tell. A full AI200 rack draws the same envelope as a high-end GPU rack, which Qualcomm itself calls unprecedented density for an inference box, so the efficiency claim actually lives one level down, at performance per dollar per watt, throughput per megawatt, not absolute draw. That buys a power-starved European site more served tokens from the same grid connection, which helps at the margin and does not dissolve the electron gap by one volt. The real efficiency leap is the AI250’s near-memory design, and that is a 2027 product, unproven at scale, landing in the exact window the gigafactories are meant to be choosing winners and pouring foundations.
The case for putting Qualcomm in a European build is stronger than the sovereignty purists will grant. A memory-heavy inference node serves a very large model from fewer, cheaper boxes, which is the TCO math that matters when ten billion euros of public money is being asked to do the work of thirty, the same de-risking pattern I traced in my data-center incentives piece. A credible third inference supplier leans on Nvidia’s pricing, and the tender explicitly scores bids on avoiding supplier lock-in, the one criterion where Qualcomm’s open-ecosystem, one-click-Hugging-Face story lines up with something Brussels is grading. Confidential computing sits in the racks already, which fits the trustworthy-AI language the Commission drapes over everything. On paper the pitch is coherent.
The case against is heavier, and I lean that way. Inference-only means Qualcomm structurally cannot anchor a training gigafactory, so its role depends on those campuses being built as train-and-serve sites rather than pure training farms, a design choice nobody has committed to. The track record is thin: Centriq got killed, Cloud AI 100 never won broad deployment, and the only at-scale customer Qualcomm can name is Humain, the Saudi state-backed outfit taking 200 megawatts of this gear from this year. One flagship customer, state-owned and in a single country, is not the deployment history a European procurement officer wants under a decade-long bet. Nvidia’s real moat was never the transistors, it was CUDA, and Qualcomm’s software answer has never been stress-tested against CUDA at frontier scale, so calling it open is a positioning choice rather than a proven moat. And the CPU half of the pitch rides on NVLink Fusion, which means Qualcomm reaches the GPU through Nvidia’s own interconnect. Sovereignty by way of the incumbent’s fabric is a peculiar sort of sovereignty.
The letter that commits to nothing
What the letter of intent actually commits to is nothing. It is not a purchase order. It reassures bidders that if they win, the hardware will be there, and the Commission signed three of them, AMD and Nvidia included, mostly to de-risk the optics of a tender that runs entirely on American silicon. Those LOIs were signed following the EU-US trade agreement, and that is the detail I cannot get past: the hardware access under Europe’s flagship sovereignty project is a deliverable of a trade deal with the country it is trying to be sovereign from. The European answer, the SiPearl CPU inside the ÆTHER bid I flagged when I wrote about who controls the compute, the models, and the off switch, is the one that has to ship for any of the sovereignty talk to hold, and it is nowhere near Qualcomm’s readiness.
A couple of things I could not find, and the gaps say something. There are no independent inference benchmarks at gigafactory scale, so the perf-per-watt figure is a spec sheet, not a result. Nobody has explained where 768 GB of LPDDR per card comes from at gigafactory volume while the memory market is already in a crunch. And I am not touching the Alphawave SerDes and NVLink Fusion plumbing here, that is its own teardown for another day.
I went in expecting to wave off the low-power pitch as vendor noise, and I came out half-convinced on one narrow axis and more skeptical on the rest. Qualcomm has quietly built a real data-center stack across four years and a decade of scar tissue, and its instinct that inference efficiency is where the European power math breaks is correct. But an inference specialist volunteering to help build training machines, on American silicon, routed through Nvidia’s fabric, backed by one Saudi customer and a non-binding letter, is not the sovereign-infrastructure story Amon’s paragraph wants it to be. Whether any of it lands in a winning bid comes down to a design question nobody in Brussels has answered: are these campuses meant to serve models as much as train them, or are they pure training farms? If it is the latter, the letter of intent is a press release with a signature on it. The Commission keeps saying training and Qualcomm keeps saying inference, so you can guess which way I am leaning.