Newsletter · · Ashutosh Agarwal

Broadcom Lands Apple's AI Server Chip Through 2031 - Custom Silicon vs Nvidia - Week of July 13, 2026

Custom Silicon vs Nvidia for the week of July 6–13, 2026. Apple's chip rumor becomes a signed Broadcom design win through 2031, Meta puts September on its own Iris accelerator, and a record SK Hynix listing turns memory into the trade of the week.

Custom Silicon vs Nvidia

Week of July 13, 2026: Broadcom Lands Apple's AI Server Chip Through 2031

A note before we start: Be honest about the mix. Two things dominated: a genuine design-win (Broadcom officially landing Apple's first AI server chip) and a memory mania (SK Hynix's blockbuster US listing, Samsung out-earning Nvidia). The custom-silicon-vs-Nvidia debate itself moved mostly through pundits and reporters this week; the one real operator on camera was a memory chairman, not a hyperscaler chip boss. So: loud on the picks-and-shovels, quieter on the front-line ASIC programs, with one big exception, and it's the lead item.

TL;DR

  • Apple's chip flirtation with Broadcom is now official. A multi-year agreement through 2031, a ~$30 billion commitment, and, per Bloomberg's Mark Gurman, a first-ever Apple server chip codenamed Baltra, "as early as next year," with "4x the performance" of today's Apple Intelligence servers. Last week Apple was a rumor. This week it's a signed Broadcom design win. Broadcom is now on track for "60% of the custom AI server market by next year."

  • Meta put a date on its own chip. An internal memo (via Reuters) says Meta's in-house accelerator, codenamed Iris, goes into production this September, "a breakthrough for an in-house effort that's struggled for more than five years." Meta is also renting out spare compute and spending up to $145 billion on AI this year.

  • Nvidia is now cheaper than Hershey's. The stock has shed roughly $1 trillion in under two months and trades around 18x forward earnings, below the S&P 500 average, even as it still holds ~97% of the server-GPU market. The rotation went straight into memory.

  • Memory is the trade of the week. SK Hynix pulled off the largest foreign IPO in US history ($26.5 billion), Samsung out-earned Nvidia on profit, and the SK chairman told Bloomberg his customers don't want him to double capacity, they want "five times, six times."

What's new

Ranked by what actually moves a book, a confirmed design win first, a dated deployment next, then valuation and the memory complex, with pundit framing last.

1. Apple + Broadcom, made official: the "Baltra" server chip. [NEWS / DESIGN WIN] This is the one that matters. Last week Apple was the eighth rumored name to go custom; this week it's a signed Broadcom win. On Bloomberg Tech (2026-07-06), Bloomberg's Apple editor Mark Gurman (reporter, citing his own sourcing) laid out what the deliberately vague press release won't: "this is specifically for a new server chip that Apple is developing... This is called ASIC chips... single-purpose silicon... which is processing for AI. And so they're going to be rolling out this chip as early as next year. It's called Baltra within Apple. It's their first server-specific chip." Today Apple's Intelligence servers run on repurposed M2 Mac processors; Baltra will be "a version of the M5 Ultra... this is going to have 4x that performance on the GPU and on the CPU side." Broadcom, Gurman noted, had been "designed out" of Apple's Wi-Fi/Bluetooth stack (Apple's own N1 chip), "Now they're coming back in for AI and for server purposes." The stock popped ~4-6%, its best day since February.

The scale showed up on Squawk on the Street (2026-07-08), where CNBC's Mackenzie Segalas (reporter) pinned the number: Apple is "committing $30 billion to Broadcom in its largest U.S. manufacturing commitment to date," a figure "actually expected to top $30 billion" and tied to Broadcom "producing more than 15 billion U.S.-made chips for Apple." Apple is also putting $1.5 billion directly into Broadcom's Fort Collins, Colorado site. The read-through she gave is the one to write down: "Broadcom has become this really key stakeholder for the hyperscalers. They're on track to take 60% of this custom AI server market by next year. That's according to Counterpoint Research." Broadcom's market cap has added "$700 billion" in a year, roughly double Nvidia's one-year return. On The Rundown (2026-07-06), host Zaid Admani put it plainly: Broadcom "makes custom chips for Google, now Apple, Meta, and OpenAI are also in the mix too. So whoever wins this AI race, Broadcom will get paid." The Morning Market Briefing (2026-07-06) framed it as the headline it is: "Broadcom is back in the ASIC business," via a new long-term agreement running through 2031.

2. Meta's "Iris" chip gets a production date: September. [NEWS / LEAK] The other genuine scoreboard move. On The Rundown's weekend deep dive (2026-07-11), Admani cited an internal memo reviewed by Reuters showing "Meta plans to put its own custom AI chips, codename Iris, into production by September... having their own chip could mean less reliance on NVIDIA, which could mean cheaper computing costs for Meta and fatter profit margins on every GPU they rent out." Reuters World News (2026-07-10) added the sobering context: Iris is "a breakthrough for an in-house effort that's struggled for more than five years", the MTIA lineage finally reaching volume, and part of a plan to "spend up to $145 billion on AI infrastructure this year alone." Why it matters for a book: the same episode revealed Meta is standing up Meta Compute, a cloud business to rent out spare GPUs (Meta stock jumped 9% the day that news broke on July 1; CoreWeave and Nebius fell), and a cheaper chip makes that rental margin fatter. On Squawk on the Street (2026-07-09), the hosts tied it together with "Broadcom working with Meta" and Meta "looking to double its overall computing capacity over time." Note the caveat: "starting production in September" is a manufacturing milestone, not a deployment-at-scale number.

3. Nvidia, treated like a value stock. [PUNDIT] The rotation out of Nvidia and into everything else got quantified. On The Rundown (2026-07-09), Admani: Nvidia "has quietly lost about $1 trillion in market cap in the last 2 months... trading at its cheapest valuation since early 2019... around 18 times forward earnings," cheaper than the S&P 500 itself, "freaking Hershey's has a higher forward PE than Nvidia right now." The kicker: "The SOX Semiconductor Index is up 74% this year, while Nvidia is up just 5%." He kept the bull anchor intact, though, Nvidia still had "97% of the server GPU market as of the end of 2025," SpaceX trained Grok 4.5 on "tens of thousands of Nvidia's newest GB300 chips," and the average price target is ~$300 (≈50% upside). On Bloomberg Intelligence (2026-07-08), Mandeep Singh (analyst) explained the fear without endorsing the panic: hyperscalers "are focused more on the inference side... they've left the training market to Nvidia," but "the timeline for this is all 2, 3 years out. There will be nothing imminent." What's inspiring everyone, he said, is Google: "Google has released 8 versions of their own chip and the Anthropic models, which are the best-in-class models, were trained on Google TPUs."

4. SK Hynix's record US debut, and a chairman who won't sit down. [OPERATOR + NEWS] The memory story became the market's obsession. SK Hynix priced the largest foreign IPO in US history, $26.5 billion, second-largest share offering after SpaceX, with ADRs at $149 opening around $170 (CNBC Fast Money, 2026-07-10). The rare genuine operator on camera was SK Group Chairman Chey Tae-won, on a Bloomberg Tech special edition (2026-07-10): "we're going to double up our whole capacity" in five years, but "my customers said that's not enough. We need it more... they want us that we're closing up five times, six times." His demand view is structural, not cyclical: "that demand will be exponentially grow in the future... our supply capacity is never going to catch up," lasting "until our human society has said there was some settlement with AGI." Crucially, the long-term contracts locking this in were "not suggested by us... they actually require that", customers proposing multi-year deals to secure supply, which is exactly what turns memory from boom-bust into an annuity. He also flagged a competitive tell: SK Hynix works "with TSMC on the base die, which the others do not do," and "we don't actually compete with any customers." (One caveat worth carrying: SK is also plowing ~$1 trillion over ten years into its own AI data centers, targeting 15 GW in Korea plus 5 GW abroad, the memory maker is becoming a power buyer.)

5. The demand numbers behind the memory mania. [PUNDIT] On Fast Money (2026-07-10), two figures are worth stealing: "DigiTimes reports high bandwidth memory prices could more than double in 2027 thanks to NVIDIA's latest platform, the Rubin," and "Bank of America estimates anywhere between 35% to 40%" of the projected "$1.5 trillion" big-tech AI infrastructure spend in 2027 "goes directly to memory itself." On Halftime Report (2026-07-10), CNBC's Christina Partsinevelos framed Hynix as "bigger, cheaper, closer to NVIDIA than a Micron," pegging Hynix at "roughly 56-58% of the high-bandwidth memory" market vs. Samsung and Micron "just around 21%" each, with Micron "sold out through 2026" and HBM potentially "sold out further in through 2027." On Limitless (2026-07-09), hosts Josh and Ejaz laid out the margins: Samsung's Q2 profit of "$58.5 billion" beat Nvidia's "$53 billion," gross margins run "52%" at Samsung and "72%" at SK Hynix (vs. ~30% for Apple hardware), and, the fact every ASIC analyst should hold onto, "one gigabyte of HBM consumes the factory capacity of roughly four gigabytes of regular DRAM." Every new accelerator, GPU or ASIC, pulls the same scarce wafers.

6. The custom-silicon plumbing keeps filling in. [PUNDIT] Last Week in AI ran two useful episodes. In #250 (2026-07-07), the hosts detailed OpenAI's Broadcom-built Jalapeño chip: built on "TSMC three nanometer... the same node as Apple's latest silicon and the NVIDIA Blackwell," meaning "OpenAI is going to have to directly compete with NVIDIA and with Apple for allocation from TSMC", and "Microsoft's own Maia inference chip is also going to be on three nanometer." Same episode: Amazon is in talks to sell its Trainium chips directly to external data-center operators to undercut Nvidia, and its custom-silicon business (Trainium + Graviton + Nitro) "crossed $20 billion of annual revenue run rate in the first quarter," against AWS operating income of roughly "$14 billion." Also: Micron is investing in Anthropic and signed a supply deal, Micron's "entire calendar year 2026 is just gone for HBM supply... that includes HBM4", with the investment tied to Anthropic's $65 billion Series H. In #251 (2026-07-09), the LongCat detail is a genuine ASIC data-point: a 1.6-trillion-parameter Chinese model whose "full training run and the large-scale deployment are built entirely on AI ASIC superpods" (likely Huawei Ascend 910C CloudMatrix 384). Custom silicon is training frontier-class models today, just not in the US.

7. Etched, the merchant challenger, pivots. [PUNDIT/OPERATOR-adjacent] Limitless (2026-07-07) and Last Week in AI #251 both flagged that Etched, the transformer-only inference startup, has quietly walked back its single-architecture bet, now claiming it can run "DeepSeek, Qwen, Mamba, Llama, like all these MoE models." It has "raised $800 million across four financings," claims "over $1 billion in customer contracts," completed a first-pass A0 tape-out on TSMC's N4P (4nm) node "in under three years from seed," and is shipping first racks this summer at a reported ~$5 billion valuation, despite not having shipped a unit. Backers skew quant (Jane Street, HRT, Jump, Two Sigma) plus Venture Tech Alliance, which carries a TSMC-allocation relationship. Interesting, but self-reported: treat as a watch-item, not a thesis.

The debate

For ASICs (share and margin come off Nvidia). The cast list keeps growing and, this week, hardened: Apple went from rumor to a signed, dated Broadcom program (Baltra, next year); Meta put September on its own chip; Amazon is about to sell Trainium externally; OpenAI's Jalapeño and Microsoft's Maia are both racing for TSMC 3nm allocation. Broadcom is on track for 60% of the custom AI server market next year, and, as SemiAnalysis's Jeremie Eliahou Ontiveros argued on Catalyst with Shayle Kann (2026-07-09), "Google struck first... very large scale" in the silicon wars, financially "backstopping" Anthropic's data centers ("$50 billion of obligations... solely in the goal of supporting Anthropic's data center buildouts") specifically so Anthropic buys TPUs at "$20 billion a gigawatt." His blunt line: this is becoming "a pretty existential risk for Nvidia... you would see Trainium take more share, TPUs take more share, Nvidia share go down."

For the merchant (Nvidia keeps the pie). Every quantified voice this week kept the timeline honest. Mandeep Singh: custom chips are "2, 3 years out. There will be nothing imminent." Nvidia still owns ~97% of server GPUs and just trained the frontier (Grok 4.5) on GB300. Apple's own chip is a server accelerator for Siri, not a merchant-GPU killer; Meta's Iris is a first production run, not volume; Etched is pre-shipment with self-graded benchmarks. And the real bottleneck this week wasn't logic at all, it was memory. HBM is sold out through 2026 and into 2027, prices could double, and every ASIC competes for the exact same scarce wafers (4:1 vs. DRAM) as the GPUs. You can design a beautiful ASIC and still not be able to feed it.

Where I come out this week: the ASIC story finally added a hard fact, a rumored buyer (Apple) became a signed Broadcom design win, and Meta stamped a production month, so the cast list stopped being just talk. But the scoreboard that would actually break the merchant thesis (a published benchmark beating Blackwell, real Trainium/Maia/MTIA volume) still hasn't printed, and Nvidia's weakness is a valuation reset, not a share loss. Stay long the arms dealers who get paid no matter who wins, Broadcom above all, and the memory toll booth, and treat Nvidia at 18x as multiple compression rather than a broken thesis.

Stocks in play

  • AVGO (Broadcom), the week's biggest winner. [OPERATOR-adjacent read] Bull: Apple is now a signed, named custom-silicon customer through 2031 (Baltra, ~$30B, 15B+ chips), on top of Google, Meta, and OpenAI, and Counterpoint sees 60% of the custom AI server market flowing to Broadcom by next year (Squawk on the Street). Added ~$700B of market cap in a year. Bear: expectations are now steep; Baltra revenue is "as early as next year," not this year; any Apple in-sourcing risk remains (they designed Broadcom out once before). Watch: first Baltra volume/timeline confirmation, and whether OpenAI's Jalapeño ramps on Broadcom at scale.

  • NVDA (Nvidia). [PUNDIT] Bull: ~97% server-GPU share, still training the frontier (Grok 4.5 on GB300), ~$300 average price target implying ~50% upside; at ~18x forward it's cheaper than the S&P (The Rundown). Bear: ~$1T of market cap gone in two months; every hyperscaler and lab is building around it on inference; SemiAnalysis calls the neocloud-financing gap an "existential" strategic risk (Catalyst); reports (disputed) of its next-gen rack running a year late. Watch: the "pretty big announcements" Ontiveros expects in H2 2026 as Nvidia leans into financing neoclouds; any Rubin timing slip.

  • GOOGL (Google/Alphabet). [PUNDIT] Bull: the strongest ASIC franchise on the board, "8 versions" of TPU, Anthropic's best models trained on them, and a backstop model that turns ~$50B of balance-sheet support into ~$100B of TPU revenue (Catalyst). Bear: the same capital commitments are real obligations; TPUs still lean on Broadcom (and MediaTek) for design. Watch: TPU v7 external adoption and how aggressively Google keeps financing Anthropic's 10+ GW 2027 roadmap.

  • META. [NEWS] Bull: Iris chip to production in September plus a new compute-rental business, a credible path to monetize the $145B capex and cut Nvidia dependence; stock +~20% since July 1 (The Rundown). Bear: Iris is a first production run after a five-year in-house struggle; free cash flow has thinned as capex balloons. Watch: July 29 earnings and the next capex number; any Iris deployment scale.

  • AMZN (Amazon). [PUNDIT] Bull: custom-silicon run-rate crossed $20B (Trainium + Graviton + Nitro) against ~$14B AWS operating income, and Amazon may start selling Trainium externally to undercut Nvidia (Last Week in AI #250). Bear: Trainium's competitiveness still rides on the Anthropic co-design relationship; no fresh Rainier volume disclosed. Watch: whether external Trainium sales actually materialize, and any Trainium 3 metric.

  • MSFT (Microsoft). [PUNDIT] Bull: Maia inference chip is on TSMC 3nm alongside Blackwell and Apple (Last Week in AI #250). Bear: being a TSMC 3nm competitor to Nvidia/Apple is a hard allocation fight, and there's still no Maia 2 volume disclosed. Watch: Maia 2 timing and node allocation.

  • Memory, SK Hynix / Samsung / MU (Micron). [OPERATOR + PUNDIT] Bull: the clearest scarcity trade in AI, HBM sold out into 2027, prices possibly doubling, 35-40% of 2027 AI capex flowing to memory, and the SK chairman insisting demand outruns even a 5-6x capacity build (Bloomberg Tech). Samsung out-earned Nvidia on profit; Micron sold out through 2026 with Anthropic locking supply (Last Week in AI #250). Bear: it's still cyclical, Fast Money's Guy Adami flagged "IPO euphoria" and typical post-IPO drawdowns "more than 50%," and leveraged SK Hynix products launching next week could whip the price (Fast Money). Watch: whether long-term contracts really break the cycle, and any sign of consumer-DRAM capacity flexing back.

  • Etched (private). [WATCH] Pre-revenue at ~$5B, first-pass 4nm tape-out, architecture pivot away from transformer-only. Not investable, but a real signal on how fast a young team can reach silicon.

  • MRVL / ALAB / ARM. [COVERAGE GAP] No fresh Marvell ASIC-roadmap, Astera Labs connectivity, or Arm/Neoverse royalty color surfaced this week. That's a quiet week, not a negative, the connectivity and IP attach shows up as the new Apple/OpenAI/Meta designs land in '27.

Read-throughs

  • HBM, SK Hynix / Samsung / Micron. The dominant read-through of the week. HBM is sold out through 2026 and into 2027; prices could "more than double in 2027" on Nvidia's Rubin; and Bank of America pegs 35-40% of the projected $1.5T 2027 AI capex as flowing straight to memory (Fast Money). The structural point for our beat: "one gigabyte of HBM consumes the factory capacity of roughly four gigabytes of regular DRAM" (Limitless), every new ASIC pulls this harder, and it's a shared chokepoint with GPUs.

  • TSMC / advanced packaging (the real bottleneck). OpenAI's Jalapeño, Apple's silicon, Nvidia's Blackwell, and Microsoft's Maia are now all fighting for the same TSMC 3nm allocation (Last Week in AI #250); Etched's edge partly rests on a Venture Tech Alliance line into TSMC capacity (Last Week in AI #251). On Halftime (2026-07-10), the line was blunt: "all roads lead to Taiwan Semi. So whether you're Meta... or Apple... or Nvidia... you need capacity, Taiwan Semi."

  • Power as the new gating factor. SemiAnalysis's Ontiveros (Catalyst) reframed the whole race around electricity: "power is revenue" for the AI labs, Anthropic wants to go from 1.5 GW (end-2025) to 10+ GW by 2027, "building a Google in 2 years", and Google's willingness to backstop that power/capex is why Anthropic can commit to TPUs. He also floated a contrarian call: gas-turbine orders "are going to peak this year." Watch this as the upstream constraint on how fast any silicon, merchant or custom, can actually be deployed.

  • Custom-silicon IP / connectivity / ASIC services (Arm, Marvell, Astera Labs, Alchip). No fresh color this week, a coverage gap, not a downgrade. The attach rates matter when the Baltra/Jalapeño/Iris/TPU-v7 wave lands in '27.

What changed vs last week

Last week (July 6) the story was a list lengthening: Anthropic became the eighth name merely in talks to go custom (via Samsung), Google's three parallel TPU programs got quantified, and the freshest hardware was OpenAI's Jalapeño. This week the story hardened into facts and rotated into memory:

  • Apple crossed from rumor to signed. The single biggest delta: last week Apple was a whispered future custom-silicon name; this week it's an official multi-year Broadcom agreement through 2031, a ~$30B commitment, and a named chip (Baltra) shipping "as early as next year." A rumored buyer became a confirmed Broadcom design win, exactly the kind of hard datapoint the ASIC bulls have been missing.

  • Meta put a date on it. Last week we flagged "still zero fresh Maia/MTIA/Trainium volume." This week Meta's own chip (Iris) got a production month, September, plus a compute-rental business that gives the capex a payback story.

  • Nvidia sentiment went from "dead money" to "value stock." Last week it was a $25B bond raise into a flat stock. This week it's a full ~$1T slide to ~18x forward, below the S&P, below Hershey's, even as its 97% share holds. The reframe is complete: the market is repricing the multiple, not the franchise.

  • Memory became THE trade. Last week memory was the quiet toll booth. This week it detonated into the headline: SK Hynix's record $26.5B US listing, Samsung out-earning Nvidia on profit, a chairman saying doubling capacity isn't enough, and HBM sold out into 2027.

Still missing, as ever: a published ASIC benchmark that actually beats Blackwell, and hard shipment/volume numbers for Maia, MTIA/Iris, or Trainium. The cast and the contracts grew; the scoreboard still hasn't.