Sources: OpenAI has talked to chip designers including Broadcom about developing an AI server chip and hired former members of a Google unit that produces TPUs
looks like the real Oracle is in Congress! @modestproposal1 : It's a good article from @theinformation on Altman wanting to set up companies to design and purchase huge amounts of chips and build large scale datacenters. I also found this part interesting. [image]
Context & Ripple Effects
This report advances OpenAI’s earlier exploration of building its own chips from evaluating options toward engaging a major chip designer and recruiting specialized TPU talent. It also fits Altman’s parallel discussions with Congress over expanding advanced-chip supply and fab capacity.
Later coverage makes the Broadcom thread consequential: OpenAI was reported to be preparing an internally used co-designed chip, while also diversifying compute across suppliers. This early outreach is therefore part of a longer effort to secure more control over the infrastructure behind its models.
First-order effects
- OpenAI gains access to chip-design expertise and personnel with experience in Google’s TPU organization, potentially strengthening its ability to specify a server chip around its own workloads.
- Broadcom becomes a prospective partner in OpenAI’s push beyond off-the-shelf AI accelerators, while Google loses experienced staff from a strategically important hardware unit.
Second-order effects
- A credible custom-chip program would give OpenAI more leverage in negotiating capacity and product roadmaps with incumbent accelerator suppliers, even before a chip reaches production.
- The hiring contest broadens from model researchers to silicon, systems, and data-center engineers; Google and other AI infrastructure builders may need to defend specialized hardware talent.
Third-order effects
- If major model developers continue commissioning bespoke silicon, AI compute may shift toward mixed supplier strategies for training and inference rather than dependence on one accelerator platform.
- The combination of chip design, procurement, and data-center planning points to a more vertically coordinated AI stack, with capital access and infrastructure execution becoming durable competitive constraints.
The trend: Frontier AI developers are moving from buying general-purpose accelerators toward shaping dedicated, heterogeneous compute stacks around their own workloads.