TPU Are Not Killing GPU. The AI Chip Pie Keeps Getting Bigger
Over the past few weeks, anything tied to $Alphabet-C (GOOG.US)$ TPUs has become the new market darling. Add in reports that $Meta Platforms (META.US)$ may start deploying TPUs around 2027, and a neat but scary story appeared on trading desks in the market. The story says TPUs will replace GPUs, cloud giants will walk away from $NVIDIA (NVDA.US)$ and $Advanced Micro Devices (AMD.US)$ , and only a small TPU circle will survive.
The data tells a very different story. The AI chip market looks like a fast expanding pie, not a fixed size bowl of soup.
Clouds are growing the whole AI chip budget
Cloud providers are still the largest buyers of AI chips and they are spending more, not less.
Revenue at the major cloud platforms is growing at a healthy clip and management teams are guiding to higher capital spending over the next few years. Several hyperscalers are already spending tens of billions of dollars a year on data centers and AI infrastructure, and Street models for 2026 capex have been revised up again and again.
On the supplier side, $NVIDIA (NVDA.US)$ data center revenue, $Advanced Micro Devices (AMD.US)$ data center revenue and $Broadcom (AVGO.US)$ AI semiconductor revenue are all posting very strong growth at the same time. That only happens when the total AI budget is rising. If TPUs were really eating GPUs in a one for one way, you would not see all three curves moving up together.
The key variable is not GPU versus TPU. It is how fast the overall cloud spend on AI silicon grows. Right now that line still points up.
CoWoS is a shared bottleneck, not a TPU only highway
On the supply side, high end AI chips are all hitting the same wall. Almost every top tier accelerator, GPU or TPU, depends on advanced packaging and HBM, especially $Taiwan Semiconductor (TSM.US)$ CoWoS.
CoWoS capacity is ramping but it is still tight. Industry work suggests Nvidia will continue to take the largest share. Broadcom, as the main TPU partner for Google, has a meaningful but not dominant slice. AMD and other custom designs also need space on the same lines.
In other words, TPUs and GPUs are both constrained by the same physical bottleneck. Meta buying more TPUs is real upside for Broadcom, but those units still need to fit inside the CoWoS and HBM envelope. They cannot magically displace booked GPU capacity at scale.
Street forecasts reflect this reality. For 2026, consensus expects AI related revenue at Nvidia, AMD and Broadcom all to grow strongly. There is no sign of one line going up while another collapses. The picture is three lines climbing together as the AI infrastructure budget expands.
Summary
Google TPUs and Meta deals matter. Custom accelerators will play a big role in long run cost and performance, especially for inference.
What the current numbers do not support is the idea that TPUs will suddenly erase GPU demand. Cloud revenue and capex are still rising, Gemini 3 shows scaling laws are alive, and CoWoS remains a shared bottleneck that forces all vendors to grow within the same capacity box. Forecasts for 2026 have Nvidia, AMD and Broadcom all compounding AI revenue rather than trading places.
The smarter way to frame this is simple. The AI chip market is a multi vendor arms build out inside a rapidly growing budget, not a winner take all fight between GPUs and TPUs. For investors, the focus should be on who secures advanced packaging, who builds the strongest system and software stack, and who turns this expanding pie into durable free cash flow.
Disclaimer: Moomoo Technologies Inc. is providing this content for information and educational use only.
Read more
Comment
Sign in to post a comment
Moomoo Insights
Moomoo Official Account
Decoding markets. Delivering alpha.
35K
Followers
10
Following
100K
Visitors
Follow
空崎ヒナ誕生祭2026 : TPU and GPU are fundamentally different things. Plus, NVIDIA still got CUDA and still got the fight in it to continue improving its product. So, it will very likely coexist as they both fulfill different needs.
Mrs Durian Liew 空崎ヒナ誕生祭2026 : wait till we see Alibaba and Huawei Chip. until then, NVDA rocks
空崎ヒナ誕生祭2026 : Well, as long as Nvidia continues to be competitive, it won't be an issue.
Kipbana : Good numbers but the speculators do not look at numbers. Nvidia is in negative bias now…trading by emotions is the thing now: “Nvidia is Cisco!” “Nvidia is AI Bubble!” “Sell sell sell because Meta is going to stop using GPU”![[undefined]](https://static.moomoo.com/nnq/emoji/static/image/default/default-black.png?imageMogr2/thumbnail/36x36 "undefined")
![[undefined]](https://static.moomoo.com/nnq/emoji/static/image/default/default-black.png?imageMogr2/thumbnail/36x36 "undefined")
![[undefined]](https://static.moomoo.com/nnq/emoji/static/image/default/default-black.png?imageMogr2/thumbnail/36x36 "undefined")
![[undefined]](https://static.moomoo.com/nnq/emoji/static/image/default/default-black.png?imageMogr2/thumbnail/36x36 "undefined")
Nvidia is Enron!
The Long Game : History shows platform owners always internalize their critical silicon
Apple → Intel (Death by a Thousand S-curves)
• 2020: Apple shipped 0% ARM-based Macs.
• 2023: ARM-based/M-series Macs hit 90%+ of Mac shipments.
• Intel’s Mac market share collapsed from 100% → 10% of high-end Android share.
• Amazon’s Inferentia/Trainium now handles >50% of Amazon’s internal inference workloads (per AWS 2023/24 disclosures).
Pattern:
When a hyperscaler/platform owns both workloads and distribution, they always shift toward in-house silicon to optimise TCO (total cost of ownership).
The Long Game : The “expanding pie” doesn’t prevent share shifts — history proves this too
Example: Smartphones (2012–2020) -
but supply share still migrated:
• Qualcomm: 55% → 29% share of premium SoCs
• Apple (in-house): 0% → 32%
• Samsung/HiSilicon: Cycloidal patterns but largely displaced
Expanding TAM ≠ static share.
Silicon dominance moves with ecosystem control.
The Long Game : Even if the AI TAM grows massively, history shows dominant suppliers do get displaced when:
1. A platform controls workload + distribution (Apple, Google, Meta, AWS)
2. Custom silicon meets the workload 80/20 rule
3. Capex pressure favors TCO optimisation
4. Inference becomes the main battlefield
空崎ヒナ誕生祭2026 Kipbana : "Nvidia is Enron" is an overstretch. The Enron case is just fraud.
Kipbana 空崎ヒナ誕生祭2026 : Michael Burry doesn’t care. They just keep throwing comparisons without numbers, and distort the truths as long as it creates fear
NeMoTron : those who think TPU will take over Nvdia gpu and cuda kindly clear your position so i can add more on a discount. Thank you
View more comments...