According to Tom's Hardware, citing Fubon Research estimates reported by TechNews, Google could produce 12–15 million TPUs in 2028, a figure that would meet or exceed Nvidia's projected 12.4 million data-center AI GPU shipments that year — a gap analysts say may push Google to add Intel Foundry capacity alongside TSMC.
How large is Google's 2028 TPU target, and how does it compare with Nvidia's GPU shipments?
According to Tom's Hardware, Fubon Research's checks indicate that Google plans to have 12–15 million TPUs in 2028 — a figure corroborated by TechNews, which cites the same Fubon estimate of a "1,200 萬至 1,500 萬顆" (12–15 million unit) deployment target for that year.
For comparison, Fubon estimates that Nvidia supplied 8.2 million data-center AI GPUs in 2026 and is on track to increase that figure to 12.4 million by 2028, according to both Tom's Hardware and TechNews, which independently reported the identical 820萬顆 (2026) and 1,240萬顆 (2028) figures.
| Metric | 2026 | 2028 |
|---|
| Google TPUs (Fubon est.) | — | 12–15 million |
| Nvidia data-center AI GPUs (Fubon est.) | 8.2 million | 12.4 million |
Lined up side by side, Google's low-end 2028 target (12 million) already matches Nvidia's 2028 projection (12.4 million), and Google's high end (15 million) would exceed it — the basis for the analyst's claim that Google's unit output could surpass Nvidia's in that year.
Why would Google need Intel Foundry to hit its 2028 TPU target?
Fubon Research, as quoted by Tom's Hardware, states plainly: "Although we do not have the detailed allocation yet, we think it is difficult to reach Google's target with TSMC alone, and Intel's supply is a must by 2028." In other words, the analyst frames Intel Foundry capacity not as an option but as a requirement for Google to hit the 12–15 million unit range.
This assessment follows earlier reports, cited by both Tom's Hardware and TechNews, that Intel had landed orders to make three million TPUs for Google, following months of Google testing Intel's advanced packaging technologies. TechNews frames the same figure as "超過 300 萬顆" (more than 3 million) TPU foundry orders secured by Intel after the testing period.
What does Google's TPU v9, due in 2028, require in terms of design and capacity?
Tom's Hardware reports that entering 2028, Google's TPUs will move to the v9 generation with four compute dies, and that this design shift means the chip's "capacity consumption will more than double in 2028 versus 2027." TechNews's report of Fubon's 12–15 million unit target for that same year lines up with this jump: a doubling of per-unit capacity draw, layered onto a multi-million-unit production target, is the combination Fubon points to when arguing that TSMC capacity alone would not suffice.
How do Intel's and TSMC's advanced packaging approaches differ?
According to Tom's Hardware, Intel's EMIB/EMIB-T packaging and TSMC's CoWoS-L packaging are incompatible with each other. This technical detail matters because it means Google cannot simply shift TPU volume between the two foundries using a single packaging design — a dual-sourcing strategy would require separate packaging qualification for each foundry partner.
That qualification work appears to already be under way: both Tom's Hardware and TechNews report that Google spent months testing Intel's advanced packaging technologies before the reported three-million-unit order was placed, consistent with the idea that Intel packaging (EMIB/EMIB-T) needed to be separately validated rather than dropped into an existing CoWoS-L-based flow.
Would Google outshipping Nvidia in unit volume reshape the competitive landscape?
Tom's Hardware's own analysis, distinct from the Fubon figures, argues that even if Google's unit shipments overtake Nvidia's, this "is not necessarily bad news for Nvidia," because AI accelerator demand is expanding fast enough that "Google will simply grow faster" alongside Nvidia rather than at its expense — at least until Nvidia's next generational leap.
This reading sits against the Fubon numbers already cited: Nvidia's own projected volume still grows from 8.2 million (2026) to 12.4 million (2028) even as Google's target reaches 12–15 million in the same year — both companies scaling up in parallel rather than one shrinking while the other grows.
How does Nvidia plan to maintain its position after 2028?
Tom's Hardware states that Nvidia is expected to increase AI accelerator production with its Feynman and Feynman Ultra architectures in 2029–2030, which the outlet frames as the point at which Nvidia would next expand capacity beyond its 2028 run rate.
Separately, TechNews reports a cost pressure Nvidia faces in the nearer term: Fubon's checks suggest Nvidia's HBM4 cost will rise to US$31–32 per GB in 2026, up from US$17–18 per GB for HBM3e. That roughly-doubling per-GB cost increase is not tied by the evidence to the Feynman timeline, but it is the one concrete cost data point in the pack bearing on Nvidia's economics as it heads toward the 2028–2030 period referenced above.
What this means
The numbers Fubon has put on the record put Google's 2028 TPU target (12–15 million units) at or above Nvidia's projected 2028 GPU shipments (12.4 million) — while Nvidia's own volume is still projected to grow from 8.2 million in 2026. Fubon's own explanation for why Google needs that much capacity rests on two linked facts already in its estimates: TPU v9's four-compute-die design more than doubles per-unit capacity draw between 2027 and 2028, and Fubon says TSMC alone cannot cover that combined volume-and-complexity increase, making Intel supply "a must by 2028." That dual-sourcing plan runs into the packaging incompatibility Tom's Hardware notes between Intel's EMIB/EMIB-T and TSMC's CoWoS-L — which is consistent with reports that Google spent months separately testing Intel's packaging before the three-million-unit order materialized. Meanwhile, Tom's Hardware's own view is that Google out-shipping Nvidia in units would not by itself undercut Nvidia, since Nvidia's projected volumes keep rising through 2028 and the company has further capacity expansion plans slated for 2029–2030 with Feynman and Feynman Ultra — even as its HBM4 input costs are reported to nearly double per gigabyte in 2026.