In recent weeks, OpenAI has signed blockbuster deals with AMD and Broadcom to build vast numbers of AI chips. Much of the focus has been on the financial implications, since OpenAI will need hundreds of billions of dollars to make good on its promises. As important as it is to look at the quite implausible financials, we also need to look at the broader implications for the industry. Like, the chips themselves, what that spells for the AI industry as a whole, and the added pressure on TSMC, the only chip company that can actually build this stuff.
The Deals
OpenAI’s deal with AMD will see the chip giant build out 6 gigawatts’ (GW) worth of GPUs in the next few years. The first 1 GW deployment of AMD’s Instinct MI450 silicon will start in the back end of 2026, with more to come. AMD’s CFO Jean Hu believes that the partnership will deliver “tens of billions of dollars in revenue” in future, justifying the complicated way the deal is funded.
Meanwhile, Broadcom’s deal with OpenAI will see the pair collaborate on building 10 gigawatts’ worth of AI accelerators and ethernet systems that it has designed. The latter will be crucial to speed up connections between each individual system in OpenAI’s planned data centers. Like the deal with AMD, the first deployments of these systems will begin in the back half of 2026 and is set to run through 2029.
Advertisement Advertisement
Advertisement
Phil Burr is head of product at Lumai, a British company looking to replace traditional GPUs with optical processors. He’s got 30 years experience in the chip world, including a stint as a senior director at ARM. Burr explained the nitty-gritty of OpenAI’s deals with both Broadcom and AMD, and what both mean for the wider world.
Burr first poured water on OpenAI’s claim that it would be “designing” the gear produced by Broadcom. “Broadcom has a wide portfolio of IP blocks and pre-designed parts of a chip,” he said, “it will put those together according to the specification of the customer.” He went on to say that Broadcom will essentially put together a series of blocks it has already designed to suit the specification laid down by a customer, in this case OpenAI.
Similarly, the AI accelerators Broadcom will build are geared toward more efficient running of models OpenAI has already trained and built — a process called inference in AI circles. “It can tailor the workload and reduce power, or increase performance,” said Burr, but these benefits would only work in OpenAI's favor, rather than for the wider AI industry.
I asked Burr why every company in the AI space talks about gigawatts worth of chips rather than in more simple numbers. He explained that, often, it’s because both parties don’t yet know how many chips would be required to meet those lofty goals. But you could make a reasonable guess if you knew the power draw of a specific chip divided by the overall goal, then cut that number in half, then remove an extra 10 percent. “For every watt of power you burn in the chip, you need about a watt of power to cool it as well.”
... continue reading