Google targets 1000x AI compute rise in five years with new chips, DeepMind aid
When I first saw the memo, the headline was hard to miss: Google wants to crank its AI compute up by a thousand times in the next five years. It’s not just about piling on more silicon, though. The brief makes it clear that executives are already talking about needing next-generation chips, a tighter dance between hardware and software, and models that are leaner but smarter enough to get more done per operation.
DeepMind’s research groups have been pulled in to sketch out what future models might look like and how much horsepower they’ll actually need. Vahdat, Google’s chief architect, has been sounding the alarm - we have to sprint on building out capacity or the target could slip away. It seems the plan will hinge on a mix of more efficient AI models, brand-new AI chips, and a closer hardware-software co-design effort, all backed by DeepMind’s forecasts.
I’m still not sure how quickly the pieces will fall into place, but the pressure feels very real.
Meeting that goal will require more efficient AI models, new AI chips, tighter hardware-software co-design, and support from Deepmind's research teams, which are helping Google anticipate future model capabilities and compute demands. Vahdat said the company has to race to build out compute capacity in order to meet demand. He described the race to build AI infrastructure as "the most critical and also the most expensive part" of the AI race.
Google doesn't have to outspend competitors, he said, but it will "spend a lot" to build an infrastructure that is "far more reliable, more performant and more scalable than what's available anywhere else." Google's own hardware is a major part of that strategy. Last week, the company unveiled the seventh generation of its Tensor Processing Units, codenamed "Ironwood". Google says the new TPU is nearly 30 times more energy-efficient than the first cloud TPU introduced in 2018.
OpenAI CEO Sam Altman recently made a similar point, arguing that the AI race ultimately comes down to securing as much compute as possible. To keep pace with chip makers and cloud providers, OpenAI is taking on significant debt. Even Google employees worry about a potential AI bubble At the same meeting, Google employees raised concerns about the financial risks tied to these investments.
CEO Sundar Pichai acknowledged those worries, noting that fears of an AI bubble are "definitely in the zeitgeist." Still, he argued, as before, that underinvesting would be riskier than spending too much. Pichai pointed to strong demand in Google's cloud business, which just recorded 34% annual revenue growth to more than $15 billion in the quarter. He said the numbers could have been even higher if more compute capacity had been available.
Google’s internal roadmap calls for a roughly 1,000-fold boost, which means doubling serving capacity every six months. That cadence leaves almost no room for slip-ups. The plan leans heavily on new AI chips and a tighter hardware-software co-design, while hoping that more efficient models will squeeze extra work out of each flop.
DeepMind’s research groups are supposed to feed forward-looking model forecasts straight into the hardware build-out. Vahdat’s slide frames it as a race against demand, but the sheer scale makes me wonder about supply-chain resilience and whether power grids can keep up. The document only looks five years ahead and says nothing about cost or how risks will be mitigated.
If model efficiency improves as expected, the plan might be doable; if not, the hardware rollout could outstrip what factories can actually produce. Bottom line: Google has sketched an aggressive compute agenda, but whether the mix of chips, software tweaks and DeepMind insight will really deliver a 1,000× increase is still up in the air.
Related Reading
Browse all Google news →
Business & Startups Oracle, NVIDIA deepen tie-up to speed sovereign AI and government digital shift
Business & Startups Skilling programs lag AI; students must prioritize aspiration and depth
Business & Startups EPAM and Cursor partner to scale AI coding for global enterprise customers
Related Topic Gemini 3 Pro builds screenshot-to-code app in two prompts, fixes bugs
LLMs & Generative AI
Related Topic Baidu opens multimodal AI, claims it beats GPT‑5 and Gemini, runs on one 80GB GPU
LLMs & Generative AI
Common Questions Answered
What is Google's target for AI compute growth over the next five years?
Google aims to achieve a thousand‑fold increase in AI compute within five years. This ambitious target requires doubling serving capacity every six months and relies on new AI chips, tighter hardware‑software co‑design, and more efficient models.
How is DeepMind contributing to Google's AI compute roadmap?
DeepMind's research teams are providing forward‑looking model forecasts to help Google anticipate future model capabilities and compute demands. Their insights guide the design of next‑generation chips and the overall hardware‑software integration strategy.
What role do new AI chips play in meeting Google's compute ambitions?
New AI chips are identified as a primary lever to stretch each floating‑point operation further and support the rapid scaling of serving capacity. Coupled with tighter hardware‑software co‑design, they aim to deliver the performance needed for the 1000x compute boost.
Why does Google emphasize more efficient AI models in its plan?
More efficient models are expected to extract more work from each flop, reducing the raw silicon required for the same level of performance. This efficiency helps mitigate the massive cost and infrastructure challenges of scaling compute at the projected pace.
What is the cadence Google must maintain to achieve its AI compute goal?
Google's internal roadmap calls for doubling its AI serving capacity every six months. This aggressive cadence leaves little margin for error and underscores the critical importance of hardware advances and model efficiency.