Samsung vs TSMC: OpenAI's Foundry Shift, HBM4E & the Packaging Crunch

Jamie Cole (0:00)

AI Hardware & Chips, Daily News. I'm Jamie Cole, thanks for joining me.

Today, OpenAI's Samsung Gambit, breaking TSMC's stranglehold on AI logic.

OpenAI is moving to break its dependence on TSMC for AI logic silicon, and the vehicle is a Foundry partnership with SEMSUM built on Broadcom's $200 billion manufacturing MOU. This is the clearest signal yet that the AI chip supply chain is entering a multi-Foundry era. OpenAI's first inference chip, Jalapeno, was built exclusively on TSMC's N3P process.

The Gen2 chip is now in evaluation at Samsung Foundry's S52P mode. No contract has been signed publicly, and the process node and timeline haven't been officially confirmed. But the evaluation itself is significant. It tells us OpenAI is treating TSMC access as a structural risk, not just a sourcing preference. The key enabler is yield. Samsung's S52P process has stabilized at 70% yield, which clears the threshold for mass production viability. That's not a mature node yet, but it's enough to make the calculation credible. Meanwhile, TSMC just posted August revenue of $16.35 billion, up 53% year over year. That's the fourth consecutive monthly record. And yet the constraint limiting customer growth isn't way for capacity anymore. It's packaging. TSMC's executives have described the demand-supply gap as very big. 20 simultaneous fab projects are underway. But advanced packaging, the assembly step that stacks chips into the heterogeneous systems AI accelerators require, is tighter than the fabs themselves. That's a structural shift worth watching. Wafer capacity scales in predictable cycles. Packaging capacity is messier, more specialized, and takes longer to relieve. Advanced nodes now represent 77% of TSMC's wafer revenue. 3 nanometer is at 30%, 5 nanometer at 33%, and 2 nanometer contributed 3% in its first commercial quarter. High-performance computing reached 66% of total revenue. The picture is of a company printing money while its customers queue for assembly slots. On memory, Samsung has shipped its first Hbm4E samples. Projections put Samsung's Hbm market share at roughly 40% by the fourth quarter of 2026, up from 33% in Q2. That's a measurable shift in a market SK Hynix has dominated. Here's the thing. The important distinction here is that Samsung isn't just competing on memory in isolation. Its combined foundry memory and packaging capabilities create an integration play that TSMC's pure play foundry model can't easily match. If OpenAI commits to Samsung for logic, bundling memory and packaging under one roof becomes a real structural advantage. Not everyone is chasing HBN. Positron raised $875 million in a Series C led by NEA, reaching a $5 billion valuation on a chip that deliberately avoids HBN entirely. Their Asimov chip uses Lpddr5X memory, the same technology found in smartphones, targeting TSMC 3nm tapeout at the end of 2026, with production in the second half of 2027 The bet is that bandwidth constraints are solvable with architecture, rather than expensive high bandwidth memory. The performance claims are currently based on simulation. Production silicon is a real test. Two other developments sharpen the picture. Marvell expects its XPU business to more than double in fiscal 2028, with acceleration into 2029

Microsoft's Maya 300 chip has 300,000 units planned for 2027

That growth rate is running ahead of guidance from both Nvidia and Broadcom in custom silicon, which tells you something about where hyperscaler CAPEX is flowing. And Samsung's Taylor Texas facility has a confirmed 2027 production start for SF2P plus 2nm class chips, with realistic high-volume delivery expected in 2028

That's the window OpenAI's Gen2 timeline is targeting. The signal to track from here is straightforward. Watch whether OpenAI's Samsung evaluation converts to a signed contract, and watch whether SF2P yield holds at scale when production volumes arrive. 70% is viable. Mature and consistent is a different standard.

The packaging bottleneck at TSMC, Samsung's MemoryShare trajectory and Marvell's XPU ramp are all secondary confirmation of the same underlined thesis. The infrastructure layer of AI is fragmenting, and the companies that control multiple steps in the stack are structurally better positioned than those controlling just one.

Thanks for listening. This podcast was built using AI technology, a YesWe production.