Multiple industry sources now indicate that Gemini 3 was trained entirely on Google’s own TPUs, with no Nvidia GPUs involved in the training stack.

Google has introduced its newest generation accelerator, Ironwood (TPU v7), which features:

Approximately 4,614 TFLOPs per chip
192 GB HBM memory
7.2 TB/s memory bandwidth
Scaling to 9,216-chip pods, delivering over 42 exaflops of compute capacity

Ironwood is positioned as Google’s most powerful and energy-efficient TPU, capable of both large-scale training and high-volume inference.

In parallel, Google has secured major multi-year commitments for TPU usage, including Anthropic’s plan to use up to one million TPUs by 2026, representing one of the largest AI-compute deployments ever announced.

Bottom line:
Gemini 3 isn’t just another flagship model. Its existence is direct proof that Google’s in-house TPUs are now strong enough to train cutting-edge frontier models at scale, without relying on Nvidia.

2. Why This Is a Big Deal (Moats and Margins)

a) Google Avoids the “Nvidia Tax” on Frontier Training

Across cloud providers, custom accelerators such as Trainium and Inferentia have demonstrated:

Roughly 50–70% lower training cost versus Nvidia A100/H100
Significantly lower inference cost when fully utilized

TPUs follow the same economic logic:
a custom ASIC tuned for Google workloads dramatically reduces cost per training token and cost per inference token.

If Gemini 3 was trained entirely on TPUs, then:

Google keeps much more of the compute margin
None of the training budget is transferred to Nvidia
Frontier-model training becomes hundreds of millions of dollars cheaper per generation

This is a structural advantage, not an incremental one.

b) Infrastructure Moat: Chip → System → Model

Microsoft openly describes its Maia accelerator and Cobalt CPU as a systems-level vertical stack: chip + rack + network + software + model.

Google now possesses the same — but much more mature — with:

Custom chip: TPU v7 Ironwood
Custom networking & optical switching: Google’s internal interconnect fabric
Custom frontier models: Gemini 3 and successors, tuned directly for TPU performance

This three-layer integration forms a deep competitive moat.
Once a company controls:

The chip
The system architecture
The model that runs on it

…competitors cannot easily replicate the efficiency, no matter how many GPUs they buy.

c) Cloud Differentiation & Pricing Power

Google Cloud positions Ironwood TPUs as delivering:

Roughly 4× performance of prior TPU generations
Real-world enterprise cost reductions up to ~60%

If Gemini-class models run on the same TPU infrastructure, Google can:

Offer aggressive pricing for AI API usage
Maintain higher margins than GPU-dependent rivals
Use TPU access as customer lock-in (Anthropic being the flagship example)

This strengthens Google Cloud’s competitive position against AWS and Azure.

3. How Competitors Stack Up

Google (TPU + Gemini 3)

Hardware: 7th-gen TPU (Ironwood), exascale pods, optimized for training & inference
Software: XLA, JAX, Vertex AI; full internal integration across Search, Ads, YouTube
Model: Gemini 3, marketed as state-of-the-art in multimodal and scientific reasoning
Economics: Minimal Nvidia reliance → better margins & cost control

Microsoft + OpenAI (Maia + GPUs)

Hardware:
- Massive Nvidia H100/GH200 deployments
- Maia 100 accelerator and Cobalt 100 CPU now in production
Model: o-series (o3, o1), GPT-4.x
Economics:
- Still heavily pays Nvidia
- But expected future margin uplift from shifting workloads to Maia/Cobalt

Amazon (Trainium / Inferentia)

Hardware: Trainium/Trainium2 for training; Inferentia/Inferentia2 for inference
Model: No single flagship consumer model; AWS focuses on infra + Bedrock partner models
Economics: Demonstrates up to ~50% cheaper training and ~70% cheaper inference vs Nvidia
Position: Similar thesis to Google’s — own the chip to cut Nvidia costs — but without a showpiece model like Gemini 3

Meta (MTIA)

Early-stage accelerators focused mainly on ads and recommendation systems
Not yet at frontier-scale LLM training competitiveness

Nvidia

Still the default for anyone without custom silicon
Hyperscalers continue to buy large quantities of H100/GB200
But the high-end monopoly position is weakening as Google, Microsoft, and Amazon scale their own chips
Future Nvidia pricing power likely to decline as workloads migrate to in-house accelerators

4. Is “Gemini 3 on TPUs” Truly a Game-Changer?

For Google vs Competitors

Yes.
This is a strategic turning point. It shows Google can:

Train a frontier model entirely on its own chip
Deploy that same chip commercially
Control compute costs end-to-end

That’s an infrastructure + model moat very few companies possess.

For Cloud AI Economics

Also yes.
It validates the trend led by AWS, Google, and Microsoft:

“If you’re renting Nvidia forever, your margins are capped.”

The future belongs to companies that own:

The chip
The rack
The interconnect
The model

Owning the entire stack = owning the margin.

For Nvidia

Not an immediate collapse story — but:

A gradual erosion of premium pricing
A ceiling on unit growth
A long-term shift from “must buy Nvidia” to “buy Nvidia only where custom chips don’t fit”

The more models trained on TPUs, Trainium, or Maia…
…the less central Nvidia becomes.

For Investors

Bull case for Google

Better AI margins
More control over scaling
Lower cost per training run and per inference token
Stronger moat around Gemini and Google Cloud

Bull case for AWS & Microsoft vs Nvidia

Rising adoption of custom silicon
Long-term capex savings
Higher future cloud margins

Risks

High capital requirements for custom chip design
Execution risk at massive scale
Superior infrastructure does not guarantee superior product outcomes