Key Takeaways

  • Hyperscalers and startups alike, including Google for TPU, Meta for MTIA (Jalapeno), MatX, and Etched, rely on Broadcom to handle backend physical design and foundry execution.
  • Backend design partners run platform plays; outsourcing physical design reduces architectural differentiation to what Walter Goodwin calls “changing the color of the paint.” (play from 30:35)
  • White-label backend partnerships force competing startups onto the exact same packaging queues and High Bandwidth Memory (HBM) allocations as industry giants.
  • Fractile keeps physical design in-house to build custom 3D memory architectures, stacking DRAM directly on logic dies while sourcing alternative process and memory nodes.

The Platform Trap of Outsourced Silicon

Every wave of AI hardware startups promises bespoke silicon designed to break Nvidia's monopoly. Yet underneath the press releases, most AI chip builders share the exact same engineering backbone.

Goodwin explains that the industry leans on a single intermediary to reach production: “A lot of these chips today are built with a backend partner, for example, whether it's the TPU or Jalapeno MTIA from Meta, MatX or Etched. All of those are rumored to be using Broadcom.” (play from 27:51)

Backend partners do the heavy lifting of physical design. They take an architecture, map it to silicon geometry, and manage relations with the foundry. For a startup with limited capital, handing that burden to Broadcom seems logical. It cuts years off the timeline and sidesteps the nightmare of building a specialized physical design team.

That speed comes with an invisible tax. Backend partners do not reinvent manufacturing for every client; they plug designs into their existing, standardized pipeline. As Goodwin puts it, “A lot of the sort of building blocks are dictated by the backend partner because it's a platform technology play. You get to change the color of the paint basically.” (play from 30:23)

When ten startups use the same backend partner, they end up with near-identical memory interfaces, standard HBM configurations, and identical physical layouts. They surrender their ability to build unique hardware where it matters most: the memory pipeline.

Bypassing the Packaging Line

The real cost of relying on Broadcom shows up in the supply chain. Because backend partners reuse the same standard packaging playbooks, their customers end up waiting in the exact same line for foundry allocations and advanced HBM capacity.

Fractile chose the opposite path by keeping physical design inside the company from day one. Goodwin wanted an architecture that stacks DRAM directly on top of logic dies in three dimensions. That layout bypasses the memory bandwidth bottleneck that cripples model inference on standard chips, but it cannot be bought off the shelf from Broadcom's catalog.

“The memory stacking work, for instance, is incredibly complex and painful and requires convening multiple much larger supply chain partners to get them to do something that is highly custom,” (play from 30:42) Goodwin says.

Owning physical design allowed Fractile to seek out manufacturing capacity that competitors ignore. Instead of fighting Nvidia and Google for cutting-edge packaging and standard HBM modules, Fractile works across alternative supply chains. Goodwin points out that this path let them secure “a different process node, maybe a slightly different DRAM node or DRAM generation as well that might be going into frontier HBM.” (play from 31:16)

Bypassing the standard backend partner creates a technical moat that software tweaks cannot duplicate. If you want structural differentiation in hardware, you cannot outsource the physical implementation to a partner running a mass-market template.

What to Do With This

Audit your core technical dependencies before finalizing your supply chain architecture. Identify the single third-party vendor that translates your intellectual property into production. If that vendor uses a standardized platform that your direct rivals also license, list the architectural features you are barred from touching, then calculate the capital required to bring that specific layer in-house.