faltering

It was 2013 and Jeff Dean, one of the directors of Google, he realized something along with your team: if each Android user used their new voice search option for three minutes a day, the company would have to double the number of data centers to cope with the computational load. At the time, Google was using standard CPUs and GPUs for this task, but they panicked and realized they needed to create their own chips for those tasks. This is how it was born Google’s first Tensor Processing Unit (TPU)an ASIC specifically designed to run the neural networks that powered its voice services. That grew and grew and in 2015, before the world knew it, those first TPUs accelerated Google Maps, Google Photos and Google Translate. A decade later, Google has created TPUs so powerful that they have almost unintentionally become a surprising and unexpected threat to the almighty NVIDIA. There it is nothing. Blessed panic. Google TPUs keep their promise Until now when an AI company wanted to train its models, turned to advanced NVIDIA chips. That has changed in recent times, and in fact we have seen two recent signs that certainly pose a turning point. Missing from that timeline is the last and most striking member of this family, Ironwood, presented in April 2025. Source: Google. The first is the release of Claude Opus 4.5, an exceptional modelespecially in programming tasks. Those responsible for Anthropic already they explained that this new model does not depend only on NVIDIA, but combines the power of three different proposals: that of NVIDIA, but also Amazon’s Trainium and Google’s TPUs. But it is also that Google has given the bell because your brand new AI model Gemini 3 He has been exclusively trained using the new Ironwood TPUs that were presented in April and have become a real sensation. As we said, Google started that project in 2013 and launched its first TPU in 2015, but that internal need became a blessing, because what Google I couldn’t know is that these TPUs would end up arriving at the right time: the launch of ChatGPT turned them into a fantastic opportunity to strengthen your AI infrastructure, but also to be used for training and inference of your AI models. From there we end up reaching the current Ironwood TPUs, which in their seventh generation are exceptional both in inference as in training (as its use has demonstrated for Gemini 3). Google has managed to squeeze even more out of its chips and has doubled the peak FLOPS per watt compared to its previous generation. Source: Google. The efficiency and power of these chips gives a very notable jump compared to their predecessors, and for example they achieve double FLOPS performance per watt which was achieved with Trillium chips. If we compare them with the TPU v5p of 2023, the chips manage to reach 4,614 TFLOPS, 10 times more than the 459 TFLOPS of those models from two years ago. It’s an extraordinary leap in performance (and efficiency). The key to 2025: Google now lets others use its TPUs But in the evolution of TPUs there is another differentiating element in 2025. This has been the year in which Google has stopped “being selfish” with its TPUs. Before only she could use them, but in recent months she has reached agreements with OpenAI —which also seeks make your own chips— and especially with Anthropic. The performance of Ironwood is already comparable to that of the GB200 and even the GB300 from NVIDIA. Source: SemiAnalysis. That second alliance is especially monumental as part of that outsourcing strategy. Google is not only renting capacity in its cloud, but facilitating the physical sale of hardware. The agreement covers one million TPUs: 400,000 units of its TPUv7 Ironwood sold directly through Broadcom, and 600,000 rented through Google Cloud (GCP). In a deep report in SemiAnalysis It is revealed how from a technical perspective, the TPUv7 Ironwood is a formidable competitor. The performance gap with NVIDIA is closing, and Google’s TPU is practically the same as NVIDIA’s Blackwell chip in FLOPS and memory bandwidth. However, the real advantage lies in the cost. The Total Cost of Ownership (TCO) of an Ironwood server is estimated to be 44% lower for Google than for an NVIDIA GB200 server, allowing the search giant to offer very competitive prices to clients like Anthropic. To help even more in that race, they point out in SemiAnalysis, Google has another ace up its sleeve. This is Google’s Inter-Chip Interconnect (ICI), a network architecture that allows up to 9,216 Ironwood chips to be connected using a 3D torus topology. Google also uses optical circuit switches that allow optical data to be routed without electrical conversion, reducing both latency and power consumption. This allows you to reconfigure the topology of that network on the fly to avoid (or mitigate) failures and optimize different types of parallelism. NVIDIA’s “moat” with CUDA is narrowing We have often repeated that although semiconductor manufacturers already have flashy chips —tell AMD– In fact the true strength from NVIDIA is in CUDAthe software platform that has become the de facto standard for AI developers and researchers. Google also wants to change things here. During the last few years the company tried to focus on Python libraries such as JAX either XLAbut in recent times has started prioritizing native PyTorch support —a great competitor of TensorFlow— in its TPUs. That’s crucial to making it easier for engineers and developers to start migrating to their TPUs instead of NVIDIA GPUs. Before it was possible to use PyTorch on TPUs, but it was cumbersome, as if one had to speak a language using a dictionary in real time, while for NVIDIA GPUs that was the “native” language. With XLA Google used an intermediate library as a translator to be able to use PyTorch, but that was a nightmare for developers. Native support allows Google TPUs to behave just like NVIDIA GPUs in the … Read more