Supply-chain analyst Ming-Chi Kuo said on November 20, 2026 that Google trimmed planned 2027 shipments of Broadcom-designed TPU 8i chips by about one million units, mostly hitting the TPU 8i Max / Whalefish tier, while planning more MediaTek-designed TPU 8t silicon. Investing.com's summary of Kuo's post stresses the key gap: Google has not confirmed the figures, and MediaTek offset sizes are unsettled.
Broadcom stock sold off on the headline the same session per Investing.com market copy. For AI builders, the question is whether Google's custom silicon mix in 2027 changes TPU availability for training and serving, not whether one sell-side note moves AVGO a single day.
explainx.ai covered OpenAI's Jalapeño inference chip with Broadcom, Taiwan's AI server export rules, and Gemini Neon launches on separate tracks. This post is Google TPU procurement only.
Claimed vs verified
| Claim | Status (Nov 20, 2026) |
|---|---|
| Google cut ~1M TPU 8i units from 2027 plan | Claimed — Kuo industry survey via Investing.com, Seeking Alpha |
| Cut is ~20% of prior ~5M plan | Derived from Kuo math in outlets; not from Google |
| Cuts concentrate on 8i Max / Whalefish | Claimed — Kuo |
| Google raises MediaTek TPU 8t shipments | Claimed, magnitude unknown |
| Google or Broadcom official post | Not found in sources used here |
| Total Google AI compute demand down | Not claimed by Kuo; framed as mix shift |
Why Kuo says Google is rebalancing
Kuo's reported narrative, summarized by Investing.com:
- Frontier models eat first — Google leadership has said cutting-edge development gets priority access to compute.
- Power and supply chains are tight — data-center electricity and fab capacity force tradeoffs between chip families.
- 8t may rise as 8i falls — but without finalized 8t numbers, nobody can net out 2027 FLOPs yet.
Digitimes adds a parallel thread: Google may buy HBM directly for MediaTek-partnered chips, potentially changing procurement economics versus going through Broadcom for full stacks. That piece is separate reporting, not repeated in every Kuo summary.
Wafer grid graphic for Google TPU and Broadcom custom silicon supply claims
Custom TPU silicon planning graphic for Google Cloud capacity scenarios
What TPUs mean in Google's stack
Tensor Processing Units are Google's in-house accelerators for training and serving large models. The 8i family Kuo discusses is part of the custom ASIC road map Broadcom helps build; 8t with MediaTek would mark a second design partner if Kuo's shipment shift holds.
None of this directly changes Gemini API pricing on November 20. It matters if you:
- Reserve Cloud TPU capacity years ahead.
- Benchmark Gemini against GPU clouds assuming steady TPU generational upgrades.
- Track sovereign cloud deals that mention TPU pods.
Pair this watchlist with IEA data-center spend forecasts and Samsung foundry BIS context.
How to read the AVGO move
Market reporters tied Broadcom's same-day slide to Kuo's note. That is sentiment, not verification. Broadcom's custom ASIC business serves multiple hyperscalers; one year's Google mix does not equal total revenue without ASP and co-packaged networking attached.
Until earnings calls or SEC filings mention Google TPU 8i, treat the one million unit figure as scenario planning input, not capacity you can book.
How TPU generations show up for developers
Google Cloud customers rarely order "TPU 8i" by that codename in a console dropdown. Instead, you see pod types tied to generations and topology. When Kuo speaks about 8i Max / Whalefish, read it as high-end training silicon within Google's internal roadmap, not a SKU you can compare 1:1 to an H100 reservation.
Practical impacts if the cut is real:
- Long-horizon Gemini training schedules that assumed a specific 2027 ramp may slide to 8t or GPU colo fill-ins.
- Partner chips (OpenAI Jalapeño with Broadcom) do not move in lockstep; Broadcom can still grow with non-Google ASICs even if Kuo is right about one customer mix.
- MediaTek 8t rise, if confirmed later, would be a second-source win for Google procurement, similar in spirit to multi-vendor HBM buys Digitimes described.
What would verify Kuo fastest
| Signal | Why it would matter |
|---|---|
| Google blog on TPU v6/v7 capacity | First-party shipment tone |
| Broadcom earnings custom ASIC segment | Management comment on hyperscaler mix |
| MediaTek investor day | Any Google co-design mention |
| Cloud TPU quota emails | Customers see reservation shifts before press |
Until one of those appears, keep Kuo in the scenario column and your capacity plans in the multi-cloud column.
Kuo's track record and why survey math is not a filing
Ming-Chi Kuo built credibility on Apple supply chains, then extended surveys to AI accelerators. His posts are industry checks, not earnings guidance. Investing.com and Seeking Alpha relayed his November 20 note because it moved AVGO, not because Google confirmed it.
When you model risk:
- Use Kuo to stress-test Broadcom revenue sensitivities in spreadsheets.
- Do not treat 5M to 4M units as GAAP facts.
- Pair the note with Digitimes HBM sourcing reporting on Google buying memory separately for MediaTek-partnered chips — another unconfirmed but directionally bearish Broadcom attach story.
Power and frontier-priority narrative
Kuo's reported explanation dovetails with public Gemini messaging: frontier training eats first when megawatt caps bind. Cutting Whalefish 8i Max units may free power envelopes for whatever Google classifies as 2027 frontier without admitting lower total ambition.
Data-center operators should read that as scheduling risk, not demand collapse. IEA's November outlook still charts rising grid load even when individual chip families get rebalanced.
Whalefish 8i Max: why the cut might hurt revenue more than units
Kuo reportedly concentrated cuts on TPU 8i Max, branded Whalefish in supply-chain shorthand. If ASPs on Max parts exceed base 8i silicon, a 20% unit cut can mean a larger than 20% revenue hit to the Broadcom Google line — but only if Google does not substitute 8t volume at similar margins.
We do not have public ASP tables. The useful builder takeaway is topology planning: Max-class parts often land in large synchronous training pods where scheduling slips cascade into Gemini launch windows.
When Gemini Neon and tool-retrieval APIs ship consumer features, backend teams still depend on TPU pools Google rarely markets by codename.
Broadcom investors versus builders
Stock reactions do not prove shipment cuts. They prove expectations moved. Builders care about Cloud TPU quotas and Gemini latency, not AVGO's single-day chart.
If you are a Broadcom investor, wait for earnings Q&A. If you are a Gemini fine-tuner, keep multi-region fallbacks and watch MediaTek 8t rumors for 2027 planning calls.
FAQ depth: claimed versus verified checklist
Did Google confirm the cut? No in sources used here. Kuo survey only.
Did Broadcom confirm? No in sources used here.
Will MediaTek confirm 8t offsets? Not yet; Kuo says details unsettled.
Does AVGO down 10% verify the story? No — that is market reaction.
Should I cancel TPU reservations? No without Google Cloud notice. Use Kuo for scenario planning, not purchase orders.
Where will Google admit changes? Likely earnings, Cloud Next, or research posts naming new TPU topologies — none appeared November 20.
Reading Kuo alongside other November silicon stories
The same week brought SK Hynix–Kioxia shareholder moves, Samsung foundry BIS angles, and TSMC credit-limit rumors. None of those prove Kuo's Google numbers, but together they sketch tight supply chains where hyperscalers reallocate rather than exit AI capex.
If you operate a foundation-model lab, keep two procurement narratives in mind: public cloud reservations you can see today, and 2027 custom silicon mix you cannot until Google or Broadcom speaks. Plan capacity with buffers, not single-vendor assumptions.
For Broadcom-heavy networking shops: Custom ASIC revenue is one line item; switch and optics businesses move on different cycles. A Google TPU trim does not automatically imply NIC backlog changes — avoid jumping from Kuo to entire data-center BOM cuts without evidence.
Next verification steps for builders
- Screenshot Kuo's original post if you rely on it in capacity slides; aggregator summaries lose nuance.
- Add a Google Cloud TPU quota review to Q1 2027 planning even if Kuo is wrong, because power caps are real per IEA data-center outlook.
- Track MediaTek earnings for any Google custom-silicon hint.
- Re-read Broadcom investor day transcripts for hyperscaler customer concentration language.
- Avoid retrofitting Gemini API outages to this rumor without evidence.
- Share this claimed-vs-verified framing with finance partners so AVGO moves do not rewrite your TPU roadmap slides without primary sources.
When in doubt, default to multi-cloud GPU training for near-term milestones and treat 2027 TPU mix as a watch item, not a cancelled program. Google's silence on November 20 is not proof the cuts happened; it is proof you still need primary filings before rebuilding capacity models.
Related reading
- OpenAI Jalapeño chip with Broadcom
- Gemini 4 Neon launch
- Gemini tool retrieval API
- Taiwan AI server export crime law
- NVIDIA Microsoft Ignite Vera Rubin
- IEA data-center power spend
Sources used: Investing.com — Kuo Google Broadcom summary · Seeking Alpha — Kuo headline
Shipment figures remain analyst claims until Google or Broadcom confirms them.
