Test report DSG-1465 · Rev B · tested November 17, 2026
Processors & AcceleratorsDevice under test
Google TPU Sales Projected to Hit $135B by 2028
Bloomberg Intelligence sees TPU sales climbing from $12B in 2026 to $135B by 2028, excluding cloud rentals. Nvidia's $96.2B quarter sets the bar.
- Read
- 5 min
- Words
- 1,012
- Node
- 14nm
- Operator
- Elena Vasquez
Spec summary
- Bloomberg Intelligence estimates TPU sales rising from ~$12B in 2026 to over $135B by 2028, excluding Google Cloud rentals.
- Google introduced four TPU systems in 2026 — TPU 8t, 8i, 8tx and 8i Max — versus roughly one per year previously.
- Nvidia reported $96.2B revenue for fiscal Q2 2027 (quarter ended July 26, 2026), with Data Center revenue of ~$89B, up 117% year over year.
- Oracle announced it will offer Google TPUs outside Google's infrastructure for the first time; terms undisclosed.
- Anthropic reportedly agreed to use up to 1 million TPUs in a deal valued in the tens of billions of dollars.
![Google TPU Sales Eyed at $135B by 2028 [2026]](/media/bf/bf3d27868bd531ff.1280.webp)
Bloomberg Intelligence estimates Google's TPU sales could climb from about $12 billion in 2026 to more than $135 billion by 2028 — an 11-fold jump in two years, and one that excludes chips rented through Google Cloud. The forecast, reported by Yahoo Finance on November 9, 2026, lands in a year when Google introduced four TPU systems for the first time: the TPU 8tx and TPU 8i Max joined the TPU 8t and TPU 8i.
The number is an analyst estimate, not reported revenue. That distinction runs through the whole story, and it matters for anyone comparing Google's trajectory with Nvidia's confirmed results.
What changed in Google's 2026 TPU roadmap?
Google used to ship roughly one data-center TPU per year. Four launches in 2026 signal a company treating TPUs as a product family, not an internal side project. The naming hints at segmentation — "tx" and "Max" variants now sit alongside the base "t" and "i" parts. Reports don't spell out what each suffix means, and no specifications have been published.
The segmentation logic is straightforward. A lab training a frontier model has different priorities from a trading firm running inference, and both differ from a university with a fixed budget. A broader lineup lets Google pitch each buyer a tailored option.
For Alphabet watchers, the framing has flipped. Chips used to appear in coverage as a cost-saving tool for Google's own services. Bloomberg Intelligence now models them as a sales line that could rival established silicon vendors.
How big is the Nvidia baseline?
For the quarter ended July 26, 2026, Nvidia reported $96.2 billion in revenue, up 106 percent year over year and 18 percent sequentially. Data Center revenue came in at about $89 billion, up 117 percent. Guidance for the following quarter sits near $108 billion with gross margin around 74 percent.
Two details stand out. Data Center is roughly 92 percent of Nvidia's total revenue. And the guided $108 billion implies sequential growth cooling from 18 percent to about 12 percent.
Run the arithmetic on the TPU forecast and you get about 235 percent annual growth for two straight years. Nvidia's 117 percent Data Center growth, repeated for two years, would yield roughly 4.7x. That's the mathematical heart of the "faster track" framing — with heavy caveats attached.
Faster track or smaller base?
The two growth stories sit on very different foundations. Nvidia's figures are reported results. Google's are an analyst projection that excludes cloud rentals.
Nvidia's single-quarter Data Center revenue of about $89 billion is roughly 7.4 times Bloomberg Intelligence's full-year 2026 TPU estimate. Annualize Nvidia's latest quarter and you reach around $356 billion; the $135 billion TPU forecast would equal about 38 percent of that run rate.
Percentages also flatter small bases. A business doubling from $12 billion adds $12 billion. A business at $89 billion per quarter growing 20 percent adds more than $17 billion in a single quarter. Nvidia leads in absolute dollars under any plausible reading and would keep leading through 2028.
What does the Oracle announcement change?
Oracle said it will offer Google TPUs, letting customers use them outside Google's own infrastructure for the first time. Until now, TPU capacity meant going through Google.
Three consequences follow:
- The buyer pool widens to companies already running on Oracle's cloud.
- Google gains a second sales channel that doesn't require migrating customers onto Google Cloud.
- TPUs meet Nvidia GPUs on neutral ground, inside environments buyers already trust.
Commercial terms, launch timing, capacity and pricing were not disclosed. The move only makes sense if Google believes TPU demand will outrun its own data centers, or if it sees more upside in chip volume than in platform lock-in.
Who is buying TPUs?
The biggest named commitment belongs to Anthropic, which reportedly agreed to use as many as 1 million TPUs in a deal valued in the tens of billions of dollars. Exact contract value and final quantity are unspecified, so "up to" is the correct framing.
Google Cloud CEO Thomas Kurian put the demand picture this way: "Nearly 100 customers use TPUs today, and we're seeing strong and growing demand from many others." That's a sales statement from the executive running the business, not an independent audit — but it's a useful count.
Reported customers include:
- Anthropic
- Apple
- Physical Intelligence
- Citadel Securities
- Jump Trading
- Caltech
- Cassava
- Cirrascale
The mix spans AI labs, a consumer hardware giant, trading firms, a university and cloud-adjacent providers. Kurian's statement suggests demand is broad rather than concentrated in one buyer.
What remains unconfirmed?
A fair read requires a clear list of unknowns. The reporting gives no performance figures, memory specifications, power ratings or pricing for any of the four chips. No shipment volumes, margins or recognized TPU revenue appear anywhere. Benchmark claims circulating on social media don't appear in verified reporting.
The claim that Google's TPU business is growing faster than Nvidia's AI business rests partly on projections. Any article treating analyst estimates and reported results as equivalent data points overstates the evidence.
What should buyers do now?
Teams planning AI infrastructure shouldn't treat this as a reason to switch suppliers overnight. Benchmark your own workload on both platforms. Get written pricing and delivery dates from Google, Oracle and Nvidia's channel partners. Treat claims about speed or cost savings as unproven until a vendor puts numbers in a contract.
Developers should weigh portability: frameworks that run on several accelerators reduce lock-in, and a small pilot reveals friction before a large commitment. Finance teams should model scenarios rather than point estimates — with Nvidia guiding to about 74 percent gross margin and a competitor emerging, pricing could move either way.
The clearest signals to watch: the first earnings report that breaks out TPU revenue, or a second major cloud partner following Oracle. Until then, the $135 billion figure is a directional signal from a serious research shop — not money Google has booked.
via shattered.io (Source)
More from Elena Vasquez
Show full bio
Senior reporter covering people, launches and day-to-day practice in hardware layer of artificial intelligence at Die Signal.
675 articles
Same lot · LOT-362D
- DSG-68737nmGoogle splits TPU line into agent and security systems
- DSG-85813nmBroadcom vs Alphabet: Who Hits $100 Billion in TPU Revenue First?
- DSG-890920nmGoogle Cloud Doubles to $30.71B as TPU Sales Move the Revenue Line
- DSG-385828nmGoogle Ships 4 TPU Systems in One Year: 8tx and 8i Max Break Silicon Cadence
- DSG-15445nmGoogle Splits 8th-Gen TPU Line Into Four Chips