Chipflation Rising: AMD Zen 6, HBM Price Surge & Nvidia's Gaming GPU Gap

**Jamie Cole** (0:00)
AI Hardware and Chips Daily News. I'm Jamie Cole. Thanks for joining me.
Today, AMD server first bet why Zen 6 skips consumer by 12 plus months. A&D just confirmed its launching Zen 6 exclusively on server hardware, and consumer users won't see it for at least 12 months.
That's the clearest signal yet that the data center has fully displaced the desktop as the proving ground for x86 architecture. The event is AMD's Advancing AI Conference on July 22nd and 23rd. What's launching is 6th generation EPYC Venice. Built on TSMC's 2nm process, it scales to 256 cores and delivers roughly 1.7 times the performance of its Turin predecessor.
That's a meaningful generational step. The important distinction is this, Zen 6 has been confirmed for servers and nothing else. Desktop rising users are looking at H20206 at the earliest, more likely CES 2027 and AMD hasn't officially said either way.
This breaks a 7-year pattern. From Zen 1 through Zen 5, AMD typically launched server and consumer silicon on similar timelines. Now the server business is pulling the road map. That tells us something about where the margin incentives are pointing. Part of what's driving the server-first calculus is memory economics, and they're getting worse. HBM-4 prices are forecast to roughly double between late 2026 and 2027, moving from around $2 per gigabit to $4 or $5 per gigabit. HBM-3e is rising too. Samsung, SK Hynix, and Micron between them supply about 95% of global HBM, and all three are sold out to AI hyperscalers through 2027 No new entrants are viable before 2028
The key implication is that memory has become the strategic chokepoint in this entire hardware cycle. It's not just a cost input, it's shaping which products get built and which get shelved. The clearest example of that is Nvidia. The RTX 5080 Super was designed and then not manufactured. The GD-R7 memory earmarked for it was redirected to higher magnet AI products. This is the first time in roughly 30 years that Nvidia has skipped a gaming GPU generation entirely. That's not a supply delay, that's a strategic reallocation.
Here's the thing, for AMD and Intel, a gap in Nvidia's consumer lineup is an opening. Whether either company can actually capitalize on it, especially with their own memory constraints, is the test that matters next. Meanwhile, the hyperscalers are waiting for the GPU vendors to sort this out. Meta's custom Iris accelerator, designed with Broadcom and manufactured by TSMC, is entering production in September.
Iris is built for AI inference and it's designed specifically to reduce Meta's dependence on Nvidia and AMD. Meta joins Apple and Google as a scaled operator of custom silicon. That's three of the world's largest compute buyers partially stepping off the standard GPU stack. The signal here is sustained margin pressure on the merchant GPU market and Broadcom quietly becoming one of the most important companies in AI infrastructure. There's one more development worth watching carefully, though it comes with a caveat. There are reports that Nvidia's next-generation Kyber NVL144 AI rack could be delayed to 2028 due to circuit board manufacturing issues. If that holds, it opens a meaningful window for AMD's MI series accelerators and Google's TPUs and high-end AI infrastructure. Consider this. The caveat is this, the report hasn't been independently confirmed, and Nvidia hasn't made specific denials. Treat it as an unresolved proof point, not a confirmed shift. Step back and the pattern across all of this is consistent. HBM scarcity is reshaping roadmaps. Compute costs are inflating durably. Hyperscalers are vertically integrating to escape vendor margins. And the gap between large and small players in AI infrastructure is widening with every pricing cycle. For investors and engineers tracking this layer, the near-time metrics to watch are TSMC's 2nm yield progress on EPYC Venice, the September Meta Iris production ramp and any official confirmation on Kyber's timeline. Those three data points will say more about where AI hardware is heading than any roadmap slide.
Thanks for listening. This podcast was built using AI technology, a YesOui production.