Sam Altman desires as much as $7 trillion for AI chips. The pure assets required can be ‘thoughts boggling’

Yesterday the Wall Avenue Journal reported that OpenAI CEO Sam Altman desires to boost as much as $7 trillion for a “wildly-ambitious” tech undertaking to spice up the world’s chip capability, funded by traders together with the U.A.E. — which in flip will vastly develop its capability to energy AI fashions.

Whereas this may increasingly merely be a dreamy moonshot on Altman’s half, or an Elon Musk-like hype-generator, what shouldn’t be unsure is the environmental impression of such an enormous effort, in accordance with Sasha Luccioni, local weather lead and researcher at Hugging Face.

“If it does work out, the quantity of pure assets that might be required is simply mind-boggling,” she advised VentureBeat. “Even when the power is renewable (which it isn’t assured to be), the amount of water and uncommon earth minerals required is astronomical.”

For comparability, in September 2023 Fortune reported that AI instruments fueled a 34% spike in Microsoft’s water consumption; Meta’s Llama 2 mannequin reportedly guzzled twice as a lot water as Llama 1; and a 2023 examine discovered that OpenAI’s GPT-3 coaching consumed 700,000 liters of water.

And past the environmental impression, shortages of uncommon earth minerals reminiscent of gallium and germanium have even helped inflame the world chip battle with China.

Luccioni criticized Altman for not specializing in extra environment friendly AI strategies to develop AI. As a substitute, she mentioned, “he’s taking a brute drive strategy and persons are calling it.. visionary?”

GPU entry has change into key Silicon Valley AI battle

However the reality is, Altman’s want to deal with the present GPU shortages and reshape the semiconductor panorama shouldn’t be uncommon. Final summer season, VentureBeat reported on how entry to Nvidia’s hard-to-come-by, ultra-expensive, high-performance computing H100 GPU for big language mannequin (LLM) coaching was changing into the “high gossip” of Silicon Valley.

And simply final week, Meta provided a deep dive into its AI technique in its newest earnings name, CEO Mark Zuckerberg mentioned that to construct AI “full basic intelligence” the primary key requirement is “world-class compute infrastructure.” Zuckerberg went on to repeat what he had just lately disclosed in a current Instagram Reel: that by finish of this 12 months Meta may have about 350k H100s — together with different GPUs the full might be round 600k H100 equivalents of compute.

The corporate plans to proceed investing aggressively on this space, he defined: “In an effort to construct probably the most superior clusters, we’re additionally designing novel information facilities and designing our personal customized silicon specialised for our workloads.”

Luccioni has been important about Nvidia’s transparency concerning the carbon footprint of its merchandise (that are designed by the corporate however manufactured by the Taiwan Semiconductor Manufacturing Firm: “Nvidia has but to publish any details about the environmental footprint of their manufacturing,” she mentioned, including that e-waste as a complete can also be a “large situation as a result of folks need the brand new GPUs they usually’re basically throwing out the outdated ones after a 12 months or two.”

In Nvidia’s 2023 Company Duty Report, the corporate mentioned “emissions are generated at each stage of our product lifecycle, together with manufacturing inside our provide chain. Since 2014, we’ve anticipated our key silicon manufacturing and methods contract manufacturing suppliers to report their annual power and water utilization, waste, greenhouse gasoline (GHG) emissions, and discount targets and targets by way of the RBA Environmental Survey or CDP. We additionally count on suppliers to have their GHG emissions verified by a 3rd celebration. We use this provider information to raised perceive our product manufacturing impression and allocate carbon emissions to our clients.”

Extra transparency from OpenAI is unlikely

General, Luccioni maintains that there’s much less transparency right now in terms of the environmental impression of AI — and is unlikely to vary anytime quickly with Altman’s new fundraising march.

“In case you have a look at the PaLM 1 paper from Google, which was in 2022, after which Palm 2 [released in May 2023], the quantity of knowledge they supplied drastically dropped,” she mentioned. Within the unique paper, she defined, Google shared sufficient info in order that energy-use estimates might be made.

“Now [companies] don’t even say how lengthy it took [to train], what number of chips they used, there’s completely no info supplied anymore,” she mentioned.

However general, Luccioni says she isn’t too frightened: “I feel that is only a moonshot undertaking that gained’t really pan out,” she mentioned. “However that can put [Altman] on par with Elon when it comes to outlandish tasks that entice consideration and generate hype.”