If it seems like OpenAI is shaking up the IT market every other day or so, that is because that is precisely what it is doing. The company needs dozens and dozens of gigawatts of compute capacity to pursue its AI superintelligence dreams – perhaps more – and that means setting up suppliers to compete against each other and lining up astronomical amounts of financing to back up the deals it is cutting with datacenter and system builders.
Like everyone else, we thought that Broadcom already had a deal with OpenAI, and wrote as much back in early September in Broadcom Lands Shepherding Deal For OpenAI “Titan” XPU. This is when Broadcom went over the financial results for its third quarter of fiscal 2025 on August 3, and Hock Tan, Broadcom’s chief executive officer, let it be known that in addition to the three hyperscalers and cloud builders that have contracted with Broadcom to shepherd their custom AI XPUs through the Taiwan Semiconductor Manufacturing Co foundries – that would be Google, Meta Platforms, and ByteDance – a fourth XPU customer had emerged and Broadcom had just “secured over $10 billion of orders of AI racks based on our XPUs.”
We assumed that Broadcom was citing the value of the racks to pump up the numbers and that it would only receive a piece of the $10 billion action, but have subsequently confirmed that Broadcom is indeed building AI cluster racks for this customer and will be generating all of the $10 billion for its revenue stream. (How much Broadcom is subcontracting – if at all – remains to be seen.) And with this move, Broadcom has expanded from being a chip and board maker to an original design manufacturer, or ODM. And, a pretty big one at that.
So on Monday, which was a holiday for us at The Next Platform, when we heard that Broadcom’s stock was up like a rocket thanks to its formal announcement of a deal with model builder OpenAI, we didn’t give it a second thought, kept sipping our coffee, and working on some carpentry projects.
But as the morning wore on, we caught wind of a CNBC report on Bloomberg radio whereby Charlie Kawwas, president of Broadcom’s Semiconductor Solutions Group (the chip part, not the software part), said in an interview with OpenAI president Greg Brockman that the mystery $10 billion customer was not OpenAI. If you read the response Kawwas made carefully, he said that Broadcom had not yet received a $10 billion purchase order from OpenAI, “so I hope that answers the question.” That response actually did not answer the question. But everyone took it as a backhanded confirmation that the $10 billion deal for AI cluster racks – not just chip shepherding and packaging – was not from OpenAI, and certainly not for its rumored “Titan” AI XPU for inference.
The statement put out by Broadcom said that it was working on 10 gigawatts of AI cluster gear for OpenAI that would include AI accelerators as well as Ethernet-based networking for scale up interconnects between the XPUs (and possibly the CPU hosts in the rackscale infrastructure) and across racks with scale out networks. The racks will be “scaled entirely with Ethernet and other connectivity solutions from Broadcom,” which sounds like Broadcom’s Scale Up Ethernet (SUE) over Tomahawk Ultra chips to us. Shipments are expected to start in the second half of 2026 and will be done by the end of 2029. The statement said further that Ethernet is being used for scale up and scale out in the AI racks, and that PCI-Express switches would also be used – presumably to connect to host X86 processors.
Only a week ago, when AMD inked a 6 gigawatts infrastructure deal with OpenAI and Nvidia had inked its own 10 gigawatt deal with the company, the word going around was that OpenAI was interested in 17 gigawatts of total capacity in the middle term over the next few years. With this deal, which presumably is for more compute and networking that costs less per unit of capacity and does include shepherding the rumored “Titan” XPU co-designed by OpenAI and Broadcom, the total capacity is now 26 gigawatts. In terms of gigawatts, that gives Broadcom 38.5 percent share, Nvidia 38.5 percent share as well, and AMD 23 percent share of the power draw. Presumably, AMD would have a slightly larger share of compute than Nvidia, and Broadcom would account for even more, if we assume that Titan will really drive price/performance for inference hard. We don’t know anything about the Titan chip, but it has to drive the cost per token down for inference or it would not be worth the trouble.
It is not hard to imagine the same economics in play as with Nvidia’s own “Rubin CPX” for long context window processing. The Rubin CPX uses GDDR7 frame buffer memory, not HBM, and is cheaper and not as blazingly fast. But the overall price/performance of a mixed rack of top-end “Rubin” GPUs for decoding the GenAI models and Rubin CPX for processing and maintaining the context that is added as part of a query. OpenAI is focused on driving down the cost of inference with Titan, which may be paired with an Arm-based CPU much as Nvidia does with its prior Grace-Hopper superchips, its current Grace-Blackwell superchips, and its future Vera-Rubin superchips coming next year.
It is the ODM manufacturing relationship that Broadcom seems to have with OpenAI and the mysterious Customer Number 4 announced in early September that is perplexing to us inasmuch as Broadcom is competing with its customers such as Arista Networks, which makes network gear from Broadcom chips. To be fair, any company wants one throat to choke if possible, so having Broadcom deliver racks instead of chips eliminates some of the finger-pointing between vendors when things go wrong. This also allows for Broadcom to deliver more value and to possibly drive more profits. It will certainly deliver more revenues, given that 10 gigawatts of capacity will drive somewhere around $300 billion to $325 billion in AI cluster revenues by our estimate for GB300 NVL72 systems, and that would be about 12 million Blackwell GPU chiplets.
Inference XPUs burning the same 10 gigawatts of power in the aggregate might be much more numerous and much less costly per unit, and Ethernet interconnects instead of NVSwitch like the Nvidia rackscale systems use will presumably be a lot less costly, too. Providing 1.5X to 2X the FP4 oomph for inference per dollar spent would be an interesting gap for OpenAI to open up with Nvidia iron.
Someone has to do it, and OpenAI is motivated to get it done for enlightened self-interest reasons. But even then, this will end up being a hell of a lot more money for Broadcom than that $10 billion coming from Mystery Customer Number 4. Like somewhere around $160 billion to $200 billion over four years inclusive from Actual Customer Number 5, OpenAI.