Newsletter · · Ashutosh Agarwal
Broadcom Lands Apple as Nvidia De-Rates and Memory Peaks - AI Accelerators: GPUs, Custom Silicon & Optics - Week of July 9, 2026
AI accelerators newsletter for the week of July 9, 2026 (research window June 30 to July 9). Broadcom lands Apple as its first dedicated server-chip customer while Nvidia de-rates on the same custom-silicon thesis, and the memory debate shifts from device costs to the peak-earnings question.
AI Accelerators: GPUs, Custom Silicon & Optics
Week of July 9, 2026: Broadcom Lands Apple as Nvidia De-Rates and Memory Peaks
Issue 007: Thursday, July 9, 2026
What the operators, insiders, and sharpest analysts are actually saying on the podcast tape, separated from the pundit noise. Research window: June 30 – July 9.
The lead: custom silicon's incumbents just ran the table.
For two issues the fresh operator detail came from the challengers: Groq, Etched, Qualcomm's data-center debut. This week the story snapped back to the merchant-ASIC incumbents, and the datapoint is a genuine coup: Broadcom has landed Apple.
Per Bloomberg's Mark Gurman, Apple and Broadcom extended their partnership through 2031 to build Apple's first dedicated server chip, codenamed "Baltra," for the private-cloud servers behind Apple Intelligence (Bloomberg Tech, 7/6; Bloomberg Businessweek, 7/6). Those servers today run repurposed 2023-era M2 chips; Baltra, due 2027–2028, is described as a version of the M5 Ultra with 4x the GPU and CPU performance, with Broadcom's ASIC technology inside. State the caveat plainly: the joint press release said only "multiple generations of Apple products", the AI-server read is Gurman's reporting, not company confirmation.
By 7/8 it had a price tag. Apple is committing more than $30 billion to Broadcom, its largest U.S. manufacturing commitment to date, for over 15 billion U.S.-made chips, plus $1.5B into Broadcom's Fort Collins, Colorado site (Squawk on the Street, 7/8). The strategic read from that desk: Broadcom is "on track to take 60% of this custom AI server market by next year" (citing Counterpoint Research), has added ~$700B of market cap in twelve months, and has roughly doubled Nvidia's one-year return. Treat the 60% share and the $30B as company/analyst figures, not audited.
The tell isn't Apple joining the custom-silicon club, it's who it hired to get there. Broadcom already builds Google's TPUs; it now anchors Apple's server roadmap too. Dylan Patel of SemiAnalysis put the incumbent moat in relief: Google runs three distinct TPU design programs, "a TPU with Broadcom… a different architecture than the TPU with MediaTek… [and] a third [that] is a very different architecture", spending "hundreds of billions of dollars a year of their own ASICs" (Training Data, 6/30). Every hyperscaler wants its own chip; almost none can design it alone. That funnels to Broadcom and, increasingly, Marvell.
Marvell got the Jensen halo, and a skeptic's asterisk. Jensen Huang called Marvell the next trillion-dollar company; the "Jensen bump" is real, but so is the conflict, Nvidia is an NVLink partner and holds a small equity stake, so "some of your sales are contingent on Marvell." CEO Matt Murphy's numbers: a "record quarter… of just over 2.4 billion, up 28% year over year," guiding next quarter to "2.7 billion at the midpoint, which would be 35% year over year," with the recent Celestial AI acquisition aimed at next-gen optical networking (Chip Stock Investor, 7/7). The hosts ("we're not buying it") flag free cash flow suppressed by acquisition integration and think AMD or Intel is the likelier next $1T name. Fair.
The one insider-adjacent caution came from Qualcomm's investor day, relayed on The Circuit, 6/29: management guided fiscal-29 data-center revenue to $15B ($1B this year, $5B next) on a custom-ASIC business with "two customers… multi-generational programs", but named only one, and CFO Akash "conceded that these are not signed finalized agreements." Contrast ARM, whose comparable long-range guide was finalized. The near-term dollar is locked; the FY29 headline is not. Mark it down accordingly.
The mirror image: Nvidia is de-rating while its own ecosystem gets the halo. Nvidia has shed roughly $1 trillion in market value in under two months and now trades at its cheapest since before the AI boom, a forward P/E "lowest since 2018" (Bloomberg Intelligence, 7/8; Squawk, 7/8). BI's Mandeep Singh calls it profit-taking plus an "anticipatory" market: hyperscalers ("not just Google TPUs anymore… Amazon, Microsoft, Meta") are all building inference silicon, and OpenAI and Anthropic want their own. His own hedge: nothing competitive is "imminent," the timelines are "2, 3 years out." A China H200 approval headline briefly lifted the stock intraday. The signal for us: the merchant-GPU king is being re-rated down on the same custom-silicon thesis that is re-rating Broadcom and Marvell up. And AMD is silent on substance for a fourth straight issue, a dedicated MI350/MI400/MI450/Lisa Su/ROCm sweep again returned nothing but a 7/6 price pop. Four issues of nothing from the #2 GPU vendor is itself the data.
Memory: the debate moved from your phone bill to the peak-earnings question. Issue 006 was the downstream device-cost shock; this week the tape re-centered on the supply side and valuation. Samsung guided to 18x profit growth on HBM, with "only three manufacturers… able to produce HBM4 at scale" (AI Chat, 7/7), then promptly "sparked a chip-stock sell-off" the same day (Squawk, 7/7). The bear case, cleanly stated by Motley Fool's Lou Whiteman, is that memory is "the most commoditized part" of the AI chain and cracks first if the AI-efficiency narrative bites; his colleague noted 128GB DDR5 at ~$2,900 vs ~$800 a year ago (Motley Fool Hidden Gems, 6/26). The bull rebuttal from BI: these are "the highest margins they've ever had," customers are locking in because they fear paying more later, and the take-or-pay backbone, Micron's ~$100B backlog, ~$22B of upfront deposits, ~40% of revenue under contracts through 2030 (Chip Stock Investor, 6/30), makes "peak earnings" the wrong frame. SK Hynix files to sell ~$28B of U.S. ADRs this Friday, the cleanest live read on whether the market believes the cycle is durable. All figures are company/analyst claims.
Capex: the de-rating is a return-timing argument, and it's pundit-dominated. Gene Munster reads Amazon's $25B bond and Google's raises as implying ~37% hyperscaler capex growth next year, above the ~23% consensus, and stays constructive (Closing Bell, 7/7). Citi's Heath Terry cites "29% cash-on-cash returns" and $85B of Google equity absorbed "without resistance" (Bloomberg Surveillance, 6/25). The counter, from New Edge's Cameron Dawson: guidance has ratcheted "from 30% to 90% growth for the year," and "the equity market's patience is wearing thin" as free-cash-flow deterioration shows up (6/26). Note the absence of operator voices, no hyperscaler CFO on tape defending the number. That gap is why the strategists' patience matters.
Power, re-angled to reliability. After three issues on off-grid generation and the interconnection queue, the fresh wrinkle is grid stability: NERC issued a rare alert (its third ever) that data-center clusters are "spontaneously disconnecting from the grid during minor malfunctions," with Wood Mackenzie documenting 50+ MW load swings within minutes (InvestTalk, 7/3); PJM says data centers will drive 94% of peak-load growth through 2030 against a 2-year-build-vs-7-year-plant mismatch (TED Tech, 7/3). The operator answer is still off-grid, Bloom Energy's K.R. Sridhar says his fuel cells fed 50+ MW to an Oracle Utah site "in 55 days" (20VC, 6/29).
Negative space (which is signal). Optics still has no pure-play operator on the fresh tape, the only bridge is Marvell's Celestial AI deal; everything else is pundit stock-picking, e.g. Jensen's ~$2B stakes in Lumentum and Coherent and "capacity substantially higher than the world has today" (The MoneyFlows Show, 7/2). Coherent's CEO was last heard on 6/26, a full window ago. And Nvidia's own Rubin/Blackwell ramp went quiet on the operator tape even as the stock cratered.
What I'm watching. (1) SK Hynix's Friday U.S. ADR pricing, the cleanest live vote on cycle durability. (2) Whether any Broadcom or Apple executive confirms the Baltra server-chip framing that is, for now, only Gurman's reporting. (3) The first hyperscaler CFO to defend 90%-handle capex growth on tape, the operator voice this debate is missing. (4) A fifth issue of AMD silence would stop being an accident.