Photo: "Cylon circuit close up" by oskay, BY via Openverse

Google Could Build More AI Chips Than Nvidia Ships in 2028 — And Still Not Dethrone It

A new analyst projection puts Google's planned 2028 chip output above Nvidia's entire annual GPU shipment count — and, according to the same analysis, that still might not be enough to touch Nvidia's dominant position.

According to a Fubon Research note to clients, Google is planning to have 12 to 15 million of its own TPUs in place by 2028, the year its ninth-generation TPU is expected to arrive. For comparison, Fubon estimates Nvidia supplied 8.2 million data center AI GPUs in 2026 and is on track to reach 12.4 million by 2028. Lined up against each other, those numbers mean Google could be producing more, or at least a comparable number of, AI accelerators than the company that currently defines the AI chip market. The note describes the 2028 TPU, the v9, as carrying four compute dies, and says that design pushes capacity consumption to more than double in 2028 versus 2027 — a jump Fubon attributes directly to the four-die architecture rather than to volume alone. Four large compute chiplets on a single accelerator is a substantial engineering undertaking, and the fact that Google is building toward it signals how much confidence the company is putting behind this generation.

That confidence comes with a manufacturing problem Fubon flags directly: it doubts TSMC alone can supply enough capacity to hit Google's 12–15 million target, and the note states plainly that "Intel's supply is a must by 2028." This tracks with earlier reporting that Intel had already landed orders to build 3 million TPUs for Google, following months of Google testing Intel's advanced packaging technology. That detail matters mechanically, not just commercially — Intel's EMIB/EMIB-T packaging and TSMC's CoWoS-L are incompatible, so a chiplet has to be designed around one or the other from the start. If Google is routing volume to Intel Foundry, the decision was effectively locked in earlier in the design process, not bolted on at the end.

Stepping back, the scale here is what pushes this beyond an ordinary capacity story. If Google actually deploys more AI accelerators annually than Nvidia sells to the entire market, and Google keeps buying Nvidia hardware alongside its own chips as the source suggests it likely will, Google becomes the single largest consumer of AI accelerators in the world and, eventually, the owner of its most capable AI hardware fleet — whether that capacity ends up serving Google's own products or gets extended to others is left open. Fubon frames this as underscoring a broader shift already underway among cloud providers: designing chips around their own software stacks and workloads instead of buying off-the-shelf GPUs.

Here's the part that keeps this from being a straightforward "Google overtakes Nvidia" story: Fubon's own analysis argues that surpassing Nvidia in unit shipments would not necessarily diminish Nvidia's position, because AI demand is expanding fast enough that both companies could keep growing volume at the same time — Google would simply be growing faster, at least until Nvidia's Feynman and Feynman Ultra chips arrive in 2029–2030. Fubon points to Nvidia's current AI GPUs already being sold out as one reason raw TPU volume isn't the real threat. If that framing holds, the more useful number to watch by 2028 isn't 12 million versus 15 million — it's whether Google's TPU software stack has pulled enough workloads away from CUDA to matter, since that dependency, not chip count, is what the note points to as Nvidia's actual vulnerability.