Broadcom's $50B Bet, Marvell's $90B Forecast & the Hyperscaler Power Grab

Jamie Cole (0:00)

A Hardware & Chips Daily News.

I'm Jamie Cole. Thanks for joining me. Today, the memory pricing trap. How AI locked out PC buyers until 2028

PC memory prices are up 30% year on year, and the people paying for it aren't the hyperscalers. They're consumers trying to buy a laptop. Here's what's driving that number. High bandwidth memory production consumes roughly three times the wafer capacity of standard laptop DRAM. As AI data center demand for HBM surged, the three major memory manufacturers made a straightforward margin calculation and redirected their fabs accordingly. That wasn't a supply disruption, that was a deliberate structural choice. The consequence for the consumer market is a pricing squeeze with no near-term exit.

Intel has now raised CPU prices three consecutive times, bringing the year-on-year increase to around 30%.

Gartner's projection puts it plainly, sub $500 PCs are likely gone by 2028

Trendforce analysis of a standard $900 PC bill of materials shows memory and CPU costs rising from 45% of total cost to 58%.

The math doesn't leave room for the budget segment.

The relief date, if it comes, is 2028 That's when Micron's US plants are expected online. Until then, the consumer market absorbs what the data center industry needs. While memory markets tighten, the compute architecture underneath AI is quietly shifting in a way that matters to investors tracking AMD and Intel. Agentic AI workloads run differently from inference tasks. Agents operate autonomously for hours, sometimes days. They need sustained CPU compute for orchestration and action. GPUs handle the thinking, CPUs handle the running. The result is a ratio shift. The traditional 2 to 8 CPUs per GPU in data center racks is converging toward roughly 1 to 1 Meta's mews and OpenAI's dots both run on AMD EPYC processors. That's not a coincidence of preference. It's a function of how these workloads are structured. Morgan Stanley estimates mews alone could represent 20% of AMD's 2026 chip revenue. That's a single agentic deployment, driving a fifth of a major product line's annual revenue. HPE launched its ProLiant Gen 13 server family this week, directly targeting this shift. Four new systems, including the DL525 and XD245, use AMD EPYC 9006 SP7 Venice CPUs, with up to 256 cores per system. HPE's message to enterprise customers is direct. The constraint isn't rack space anymore. It's power. That framing tells you something about where the infrastructure bottleneck is moved. Here's the thing. The financing story around OpenAI's custom chip is worth tracking carefully, with the appropriate caution.

Broadcom has arranged discussions with Apollo Global and Blackstone for more than 50 billion dollars in debt financing tied to OpenAI's Jalapeno chip program and later infrastructure. The important qualifier is that no committed funds have been confirmed. The borrower identity and repayment terms aren't disclosed. This may not close if deployment timelines slip. What matters about the structure, if it holds, is the signal.

Moving custom silicon development from venture and equity models into institutional debt markets is a significant shift. It means Apollo and Blackstone are treating AI chip infrastructure as a credit-worthy asset class.

That's the kind of institutional recognition that changes how capital flows into the sector over the next decade. Marvell's updated revenue forecast is the clearest single data point for what the shift to custom silicon means commercially. Consider this. The company raised its fiscal 2028 target to $20 billion against a consensus estimate of $18.2 billion.

It projected fiscal 2031 revenue of $70-90 billion against consensus of $46.9 billion. That's not a modest upward revision. That's a company telling the market the ecosystem is structurally larger than analysts priced in. The signal here is that the move away from a single dominant accelerator architecture toward diverse custom silicon rewards supply chain partners. Broadcom and Marvell sit in that position. Nvidia's monoculture advantage weakens as hyperscalers build proprietary ASICs. The interconnection queue for the US national grid now exceeds 2,600 gigawatts with median approval weights above five years.

That number is the reason hyperscalers stopped waiting for grid access and started building their own power plants. Neta's Sidecat campus runs 200 megawatt gas plants with zero grid connection. Microsoft signed a $7 billion 20-year deal with Chevron for 2.67 gigawatts of gas generation in Texas. Amazon extended its Talon nuclear offtake agreement to 1,920 megawatts through 2042 You see, the implication isn't just operational independence. These companies are now equity owners in power generation assets. That's vertical integration moving upstream from procurement to infrastructure ownership. The regulatory scrutiny doesn't disappear. It just shifts from grid operators to planning authorities. That's a real risk, and one that's not fully priced. One vulnerability that moved into focus this week. Micron's Taoyuan plant in Taiwan, a critical node in the HBM supply chain, saw its union secure a 99% strike authorization on October 7 No strike date has been set. Negotiations are ongoing. But a labor action at that facility, even a short one, would ripple through AI infrastructure supply chains at exactly the moment when memory is already the binding constraint. The watchpoints from here are narrow and specific. Whether Broadcom's $50 billion financing closes, whether AMD holds its price gap against Intel through N2026, or follows Intel into a second pricing wave. And whether the Micron-Taoyuan situation resolves quietly or becomes the first major labour disruption at a critical memory choke point. The PC market won't get relief until the fabs catch up. That's 2028 at the earliest. Until then, the data centre will keep setting the price. Thanks for listening. This podcast was built using AI technology. A YesWe production.