Headline & intro
Google’s latest AI chips are not a declaration of war on Nvidia – they’re something more subtle and, long term, more dangerous: a shift in who controls the economics of AI. With the new TPU 8t for training and TPU 8i for inference, Google is trying to pull developers deeper into its proprietary stack while still hugging Nvidia tightly.
In this piece we’ll look at what Google actually announced, why Nvidia’s market value won’t suddenly implode, and how this move could reshape AI infrastructure, cloud lock‑in and digital sovereignty – especially for European companies.
The news in brief
According to TechCrunch, Google Cloud used its Next conference to unveil the eighth generation of its custom AI accelerators, Tensor Processing Units. For the first time, the line is split in two: the TPU 8t chip is optimised for training large AI models, while the TPU 8i targets inference – the phase when models are actually serving user prompts.
Google claims the new TPUs deliver up to three times faster AI training versus the previous TPU generation and about 80% better performance per dollar. They can also be clustered at immense scale, with more than one million TPUs working together in a single system.
Crucially, these chips are not replacing Nvidia GPUs in Google Cloud. TechCrunch reports that Google will continue to offer Nvidia’s latest hardware, including the forthcoming Vera Rubin GPUs, and is collaborating with Nvidia on networking optimisations such as the Falcon software-defined networking stack, open-sourced via the Open Compute Project.
Why this matters
The obvious story is “Google versus Nvidia”. The real story is “cloud versus everyone else”.
These TPUs are only available inside Google Cloud. Every time Google ships a more efficient TPU generation, it widens the performance and cost gap between:
- Customers who build on its proprietary stack, and
- Everyone trying to compete using off‑the‑shelf hardware in their own data centres or on smaller clouds.
If Google can truly offer 3× faster training and 80% better performance per dollar, hyperscale economics become even more lopsided. Foundation-model startups, SaaS vendors and internal enterprise AI teams will find it harder to justify building their own GPU clusters – both technically and financially.
The split into TPU 8t (training) and 8i (inference) is equally strategic. Training and inference costs have very different profiles:
- Training is spiky, capital‑intensive and often experimental.
- Inference is ongoing, predictable and margin‑sensitive.
By optimising separately, Google can squeeze out more utilisation – and more margin – across the lifecycle of a model. Training workloads can be sold as premium capacity; inference can be sold as scalable, cost‑efficient throughput. That’s powerful pricing flexibility.
Who loses? In the short term, not Nvidia. Google is still buying Nvidia at scale and even co‑engineering networking to make Nvidia systems run faster in its cloud. The more AI demand grows, the more Nvidia sells – regardless of who designs the in‑house sidecar chips.
The real pressure falls on:
- Smaller clouds and on‑prem vendors, who can’t match custom silicon economics.
- Customers, who get cheaper AI today but risk deeper lock‑in to a single provider’s chips, APIs and tooling tomorrow.
The bigger picture
Google’s TPU 8t/8i are part of a broader wave of vertical integration in AI infrastructure.
Amazon has its Trainium and Inferentia chips. Microsoft announced its Maia AI accelerator and Cobalt CPU. Meta is building its own accelerators for recommendation and generative workloads. All of them still buy billions of dollars’ worth of Nvidia hardware.
The pattern is clear:
- Use Nvidia as the baseline that guarantees compatibility with the existing AI ecosystem (PyTorch, CUDA heritage, frameworks, third‑party models).
- Introduce in‑house silicon where the economics are most painful: large‑scale training and high‑volume inference.
- Wrap everything in proprietary services – managed training platforms, vector databases, fine‑tuning services – that minimise how much customers think about the underlying chip.
Historically, this is not new. In the 2010s, hyperscalers did the same with networking (custom switches, smart NICs) and storage (object stores that beat generic SAN vendors on cost by huge margins). Once traffic centralised in the cloud, specialised hardware followed.
For Nvidia, the danger is not a single TPU generation. It’s a gradual erosion of addressable workload per dollar spent on AI:
- Hyperscalers might still spend more cash with Nvidia overall, but
- A growing slice of AI cycles runs on their own increasingly capable chips.
Over a long enough timeline, that shifts bargaining power. Nvidia remains the ecosystem kingmaker, but the hyperscalers quietly become less dependent on any one supplier – and more able to shape pricing, packaging and roadmaps.
The fact that Google and Nvidia are co‑developing networking (Falcon) is telling. The real moat in AI infra is not just the chip; it is the system: interconnects, compilers, orchestration, reliability. That is where scale and tight integration decide who stays profitable when AI prices inevitably fall.
The European / regional angle
For European organisations, the TPU 8 announcement is a double‑edged sword.
On the one hand, cheaper and denser AI compute in Google Cloud lowers the barrier for European startups, research labs and enterprises to train and serve advanced models. That directly supports the EU’s ambitions under programmes like the Digital Europe Programme and the Chips Act: more AI innovation, faster.
On the other hand, TPUs are the ultimate form of lock‑in. They do not ship as standalone cards you can buy and put into an on‑prem cluster in Frankfurt, Paris or Ljubljana. They live only inside Google’s data centres.
This conflicts with a growing European push for digital sovereignty and data localisation. Under GDPR, the upcoming EU AI Act, the Digital Services Act and the Digital Markets Act, regulators increasingly expect transparency, portability and control. A proprietary accelerator, only accessible via a US hyperscaler’s APIs, pulls in the opposite direction.
European cloud providers – OVHcloud, Scaleway, Deutsche Telekom, smaller regional players – generally rely on standard GPUs and CPUs. They cannot match Google’s TPU economics, which risks deepening market concentration just as Brussels is trying to diversify.
At the same time, hardware‑agnostic tooling and open standards become more important for European buyers. Anything that makes it easier to move workloads between Nvidia GPUs on‑prem and TPUs in the cloud – open formats, framework support, clear documentation – will be key to staying on the right side of EU rules about switching and interoperability.
For European policymakers, TPU 8 is a reminder: regulating AI models isn’t enough. The real leverage sits in who owns the compute layer.
Looking ahead
What happens next is less about benchmarks and more about business models.
Expect Google to use TPU 8t/8i as the spearhead for aggressive AI pricing. We’ll likely see:
- Cheaper training tiers for customers who agree to use TPUs instead of GPUs.
- Bundled AI platforms where the underlying chip is invisible – you just pay per token or per request.
- Regional expansions of TPU availability to court regulated industries that demand EU‑based processing.
Watch a few signals over the next 12–24 months:
- Ecosystem support – How seamless is it to run mainstream frameworks (PyTorch, JAX, TensorFlow) on TPU 8? If porting models is painful, many teams will stay on Nvidia.
- Vendor‑neutral tooling – Do MLOps platforms, open‑source orchestrators and European HPC centres start supporting TPUs as a first‑class target, or treat them as a niche Google‑only path?
- Regulatory pressure in the EU – The more Brussels focuses on switching costs and gatekeeper behaviour under the DMA, the more Google may have to prove that using TPUs doesn’t trap customers.
There are risks for Google, too. Over‑rotating on proprietary silicon could backfire if customers demand portability or if an open hardware ecosystem (for example around RISC‑V accelerators) matures faster than expected. And if AI spending cools, custom chip programmes become a very expensive fixed cost.
In the meantime, Nvidia will continue to sell everything it can manufacture. The inflection point, if it comes, will be gradual – visible first in how much of each hyperscaler’s AI workload sits on in‑house chips versus Nvidia.
The bottom line
Google’s TPU 8t and 8i don’t dethrone Nvidia; they deepen the power of hyperscale clouds over the AI value chain. Users get faster and cheaper AI – at the price of bigger dependence on one provider’s silicon and software stack. For European companies navigating GDPR, the EU AI Act and digital sovereignty demands, that trade‑off is especially sharp.
The key question for the next few years is simple: will you optimise for the cheapest AI compute today, or for the freedom to move tomorrow?
