Newsletter · · Ashutosh Agarwal

Nvidia Embraces Custom Chips and Google Splits Its Next TPU Two Ways - Custom Silicon vs Nvidia - Week of September 7, 2026

Custom Silicon vs Nvidia for the week of September 7, 2026: Nvidia put 3.5 billion dollars into MediaTek so custom XPUs plug into its rack, Google split its next TPU between MediaTek for training and Broadcom for inference, and both chip giants guided up hugely while their stocks fell.

Custom Silicon vs Nvidia

Week of September 7, 2026: Nvidia Embraces Custom Chips and Google Splits Its Next TPU Two Ways

The seven days to today were the busiest this letter has seen. Both of the AI-chip giants got dissected (Nvidia's print finally got its podcast treatment; Broadcom reported fresh on Tuesday night and sold off Wednesday), Nvidia dropped a genuinely thesis-moving move on the custom-silicon world, and we got real design-win detail on how Google is splitting up its next TPU.

A quick reminder of what this letter is about, for new readers: an ASIC, or "custom chip", which the industry calls an XPU, is a chip a company designs for its own specific job (Google's TPU, Amazon's Trainium) instead of buying a general-purpose Nvidia GPU off the shelf. Firms like Broadcom, Marvell, and now MediaTek don't own those chips; they're the engineering shops that turn a hyperscaler's design into a real, manufacturable chip at TSMC. The whole letter tracks whether those custom chips are eating into Nvidia's dominance, and this week, the answer got a lot more interesting.

TL;DR

  • Nvidia stopped fighting custom silicon and decided to tax it. It's putting $3.5 billion into Taiwan's MediaTek (its biggest-ever non-US investment) so MediaTek's XPUs plug straight into Nvidia's networking, NVLink, and custom memory. Jensen Huang's framing, on the record: "I see XPUs in the market. It's not coming into the market. It's in the market." The read: even when a hyperscaler builds its own chip, Nvidia wants to sell the seven other chips in the rack (Bloomberg Tech).
  • Google is splitting its next TPU in two, and it's a real design-win leak. MediaTek gets the V8 training chip; Broadcom keeps the inference chip (the higher-value, more complex one). Baird's analyst pegs MediaTek's Google revenue at $1B to $2B this year, $14–16B next year, against Google TPU volumes of ~9M units next year rising to 15–16M in 2028 (The Information's TITV).
  • Both giants guided up massively, and both stocks fell. Nvidia guided to 70% revenue growth next fiscal year (up from a prior 45%) and ~$108B for the current quarter; Broadcom said AI-chip revenue will double next year with "line of sight" to double again in 2028, yet AVGO dropped ~6% and NVDA's chart has gone sideways for a year. The market is now pricing 2027 as "locked in" and treating 2028 as the "zone of uncertainty" (The AI Investor Podcast).
  • Hock Tan says Anthropic and OpenAI will become Broadcom's two biggest XPU customers, eclipsing Google. That's the AI labs, not the hyperscalers, becoming the custom-silicon buyers (Bloomberg Intelligence, relayed on the podcast rather than from the transcript).
  • The margin story quietly turned. Nvidia guided gross margin down from 75% to 72% for its January quarter and said it stays in the low 70s for several quarters, blamed mostly on memory it can no longer fully mark up (The Circuit).
  • The whole complex now rests on financing, not chips. Nvidia's own 10-Q shows three customers = 54% of revenue and 64% of receivables; a16z's Gavin Baker can't find "a single quantitative data point getting worse"; and the bears (Dan Niles, David Bahnsen, Aswath Damodaran) are all watching the same thing, whether capital markets keep funding a buildout that isn't paying for itself yet.

What's new

Ranked by what actually moves a book: design-win leaks first, then earnings color, then the pundit debate.

1. Nvidia's $3.5B MediaTek bet, the "if you can't beat the XPU, own the rack around it" move

This is the single most thesis-relevant development of the week, and it's straight from the principals. Nvidia is investing $3.5 billion in MediaTek via convertible bonds, confirmed by Bloomberg as Nvidia's biggest non-US direct investment ever (Bloomberg Tech). MediaTek is the Taiwanese equivalent of Qualcomm (the world's biggest maker of smartphone SOCs), but in the data center it's trying to become "more like Marvell or Broadcom", the shop that helps a Google or an Amazon turn a chip idea into silicon, per Bloomberg's Ian King who broke the story.

Here's why it matters. For two years Jensen Huang has been dismissive of the custom-chip threat. This week, on the record with MediaTek CEO Rick Tsai, he changed the framing entirely (Bloomberg Tech Special Edition):

"I see XPUs in the market. It's not coming into the market. It's in the market. Look, an XPU is a specialized chip. NVIDIA's GPU is a general purpose accelerator... That explains why it is that we are the most fungible and the most durable and therefore the most rentable... computer infrastructure in the world."

The mechanism is the killer detail. MediaTek's XPUs will plug into Nvidia's NVLink Fusion (its high-speed chip-to-chip fabric), its Spectrum-X switches, and, this is new, Nvidia's custom HBM memory (NVHBM), which Jensen confirmed is "a custom HBM... built on a NVIDIA custom base layer" that Nvidia is extending to its NVLink Fusion partners. Translation: even a customer who builds its own accelerator now builds it inside Nvidia's house, using Nvidia's networking, memory base die, and CPUs.

Jensen put a number on why that's so lucrative, the best operator disclosure of the week. Nvidia's total compute economics per gigawatt of data center:

  • Hopper generation: ~$18 billion/GW
  • Grace Blackwell: ~$25 billion/GW
  • Vera Rubin: ~$40 billion+/GW

He explained there are now "like seven different types of NVIDIA chips that has to go into an AI factory. An XPU would be a replacement of one of those." So Nvidia's pitch is: fine, build your one custom chip, we'll sell you the other six, plus the fabric that ties them together (Bloomberg Tech).

For MediaTek's side, Rick Tsai quantified the opportunity: his AI-chip business is "about $2 billion this year" and he's targeting "about 15% of an addressable market of $80 billion next year."

The blunt read came from The Morning Market Briefing, which called it exactly what it is, "a power play against Broadcom": Nvidia is "basically sponsoring MediaTek to make XPUs for hyperscalers, and then basically take Broadcom and Marvell's business," while locking any MediaTek customer into the Nvidia interconnect ecosystem (The Morning Market Briefing). Whether you call it circular financing (Jensen insists it isn't: "this is not circular because obviously they do their own business and we do our own business"), the strategic intent is unmistakable: co-opt the XPU wave rather than resist it.

2. Google splits its next TPU, MediaTek gets training and Broadcom keeps inference

This is the actionable design-win leak, and it advances the Google-diversification thread this letter has tracked for weeks. Baird semiconductor analyst Tristan Guerra and The Information's team laid it out (The Information's TITV), and crucially, Broadcom confirmed it on its own earnings call:

  • MediaTek gets the V8 training version of Google's TPU.
  • Broadcom gets the inferencing version, described as "actually featuring a higher level of complexity... because they're the leader in terms of... SerDes and chip-to-chip interconnects, so they still get the higher-value design."

The numbers: MediaTek's expected Google revenue is going "from $1 to $2 billion" this year (V8 ramping late this year), and "$14 to $16 billion of revenue... from Google next year." For scale, Guerra put total Google TPU volumes at "9 million units next year going to 15, 16 [million] in 28."

The nuance that matters for the bull case on Broadcom: the analyst argued Broadcom is actually "on the right side of the equation by getting the inferencing chip," because inference will keep gaining share versus training through 2028, so Broadcom is "well positioned to actually regain some share" even after ceding V8 training to MediaTek. That's the steel-man for owning AVGO into the drop.

3. Broadcom's earnings, a beat, a huge guide, and a sell-off anyway

Broadcom reported Tuesday night (Sept 2) and fell ~6% Wednesday. The figures relayed on air (Schwab Network):

  • Q3 EPS $3.32 vs $3.23 expected; revenue $29.59B vs $29.45B
  • AI semiconductor revenue $16.7B, +221% YoY, +54% QoQ; CEO Hock Tan: demand for custom AI accelerators and networking "continues to be very strong"
  • Q4 AI-semi guide accelerating to $21.7B (+236% YoY)
  • But: Q4 total revenue guide ~$34.8B was light vs $35.05B expected, and infrastructure-software revenue ($8.75B) missed. As Schwab's contributor put it, this was "basically the first time we've had one of these major AI names actually miss figures."

The forward guide was enormous: AI-chip revenue to double in 2027 and "line of sight" to double again in 2028 (Bloomberg Intelligence). And the customer disclosure is the real story for this letter. Bloomberg Intelligence relayed Hock Tan saying Broadcom expects Anthropic and OpenAI to become its two most important customers, eclipsing Google, and that "we have six customers and four of them are just simply going to be huge." That's the AI labs, not just the hyperscalers, becoming custom-silicon buyers, funded by the financing vehicles this letter has been tracking.

Why did it drop anyway? The AI Investor Podcast gave the cleanest three-part answer (The AI Investor Podcast):

  1. Margin compression as XPUs scale (custom accelerators carry lower margins than Broadcom's legacy business), plus the mix shift of "selling Google's processors, TPUs, to Anthropic, which is going to become their largest customer."
  2. The Google share loss to MediaTek.
  3. The 2027 guide being "a little bit light" while the exciting 2028 number sits in what the host calls the "zone of uncertainty", so the market "discounts this future."

The valuation framing is worth keeping: on that podcast's math, Broadcom now trades at ~10.3x fiscal-2028 earnings against ~100% growth, versus Credo at 14.6x and Snowflake at 93x. Their punchline: "if Broadcom isn't the right stock for you, I do think something like Nvidia might be a fill-in... these companies are going to trade under relatively similar dynamics."

4. Nvidia's quarter, dissected at last

The print finally got covered. The headline figures, relayed consistently across shows (The Artificial Intelligence Show): revenue $96.2B (more than double YoY), net income $59.7B, EPS $2.46 (up from $1.08), and a current-quarter guide of ~$108B. Data center was ~$89B (~92% of the total).

Two things actually moved the debate.

The 70% guide. The number that drove the stock was Nvidia essentially committing to 70% revenue growth next fiscal year, up from a prior framing nearer 45% (The Six Five). And Jensen was explicit it's a floor, not a ceiling: The Circuit's hosts noted "Jensen kept pointing out they could grow more if they could get the capacity... he's like, it would have been a lot more". The constraint is TSMC wafers, packaging, and memory, not demand (The Circuit). The Six Five framed it as a ~$700B revenue line of sight.

The margin wobble. This is the underappreciated bit for the ASIC debate. Nvidia guided gross margin down from 75% to 72% for its January (Q4) quarter, staying in the low 70s for "the next three, four quarters", and blamed it "mostly memory." The Circuit's read: hyperscalers are increasingly "going directly to the memory companies," so the high-margin memory markup Nvidia used to capture is "going around them." That's a subtle but real crack: the merchant-silicon margin premium is not infinitely durable when your customers have the buying power to disintermediate your bill of materials.

The AWS proof point cuts both ways. As part of the earnings slate, AWS committed to another 2 million Nvidia GPUs (on top of 1 million pledged at GTC five months ago), roughly 6–7 GW, ~28,000 NVL72 racks, and an estimated $80–120B of Nvidia revenue, per The Six Five's back-of-envelope. The bull point for merchant silicon: even the company with the most successful in-house chip program (Trainium) is still buying Nvidia by the millions. As the hosts put it, "no matter how many custom AI projects are going on, there's only so many wafers... so much packaging... so much memory." The custom-versus-merchant fight is, for now, being papered over by raw scarcity: everyone wins because nobody can build enough.

Also relayed: Nvidia is acquiring Hugging Face for ~$12.9B (~80x its ~$150M forward revenue), its largest-ever deal, a move into the open-model software layer (The Artificial Intelligence Show). Less relevant to silicon, but it tells you Nvidia intends to own the ecosystem above the chip too.

5. Arm's CEO on why the CPU never went away

Rene Haas gave a rare, substantive interview (No Priors) that's directly relevant to the read-through names. His core point: the accelerator hype obscured that every one of these systems still needs a CPU to orchestrate it. His metaphor: "The token factory just generates all these tokens... where are the trucks that are going to take the tokens away and give them to the users? That's what CPUs do."

That matters because the hyperscaler CPUs inside these AI factories, Nvidia's Vera and Amazon's Graviton, run on Arm. Haas called the proliferation of custom chip designs pure upside: "a big diversification of the customer base... that can be only good." He also flagged the sector-wide constraint bluntly: "I think we're going to be in this constrained environment for three to five years at least," with the bottleneck migrating from packaging to memory to data center buildout and labor, plus rising local political opposition to data centers. And on AI-designed chips: for "straightforward designs," going from idea to a fab-ready GDS2 file could get dramatically faster, a long-term wildcard for how quickly new XPUs can reach market.

The debate

The motion: do hyperscaler custom chips structurally take share and cap Nvidia's market and margins, or do CUDA, systems integration, and Nvidia's relentless pace keep merchant GPUs on top?

This week actually shifted the debate, for the first time in a while, with fresh operator data on both sides.

For the custom-silicon side (ASICs take share). The evidence got stronger. Google is now confidently triple-sourcing its TPU (Broadcom for inference, MediaTek for V8 training), the reference custom-chip champion telling you with its checkbook that it wants competition among its designers. Broadcom's own CEO says the AI labs (Anthropic, OpenAI) will become his biggest XPU customers, eclipsing Google, meaning the buyer base for custom silicon is widening from hyperscalers to frontier labs. And Nvidia's own margin guide-down (75% to 72%, "mostly memory") is the first concrete sign that merchant pricing power has a ceiling when customers get big enough to buy around it (The Circuit).

For the merchant side (Nvidia keeps the pie). Also stronger. Jensen's MediaTek move is a tacit admission that XPUs are real, but it's also a brilliant defensive play: by making it trivially easy to drop an XPU into an Nvidia rack via NVLink Fusion, Nvidia guarantees it still sells the CPU, the networking, the switches, and the memory base die around every custom chip (six of the seven chips in the factory, at $40B+/GW for Vera Rubin). And the AWS 2-million-GPU commitment proves the point that even the best in-house program can't displace merchant silicon at scale, because the binding constraint is wafers, packaging, and HBM, not architecture (The Six Five). Jensen's own line: Nvidia is "the most fungible and the most durable and therefore the most rentable."

The most useful framing came from The Morning Market Briefing's plain-English explainer (The Morning Market Briefing): Broadcom builds the "pneumatic tubes and switches" that shoot data where it needs to go for a specific model, and its inference chip genuinely beats Nvidia for that fixed job. But "it has a smaller moat because they don't have the compiler. They don't have the map of the city... They're the contractor, whereas NVIDIA is the city planner as well." And the killer lock-in problem: "once you are locked in with a certain large language model, that's great. But if you come out with a new large language model... you kind of need a new chip for that. Broadcom has to resell that every single time." That's the real structural tension: XPUs win on cost-per-job, but they re-fight for the socket every model generation, while Nvidia's fungibility means it never has to.

Where this letter comes out this week: the smartest position is still own the design-win volume growth on both sides (Broadcom for the value-added inference sockets and networking; Nvidia for the fungible six-of-seven chips it sells around every XPU), but this week genuinely nudged the balance. Nvidia's MediaTek move converts the custom-silicon threat into a toll road rather than a competitor, which is more bullish for NVDA than the sideways year suggests, while Broadcom's margin compression plus Google share loss is a real (if overdiscounted) headwind. The one thing that would flip the view bearish on the whole complex isn't the silicon fight at all, it's the financing (below).

Stocks in play

Every ticker with fresh podcast commentary this week, with bull, bear, and the next thing to watch.

  • NVDA (Nvidia). Bull: 70% next-year growth guide (a floor, supply-constrained), ~$700B revenue line of sight, and the MediaTek move that turns XPUs into an Nvidia toll road at $40B+/GW for Vera Rubin (Bloomberg Tech). Gavin Baker "can't find a single quantitative data point getting worse" (The a16z Show). Bear: gross margin guided 75% to 72% on memory it can't fully mark up; the stock's been sideways for a year despite new highs; three customers = 54% of revenue. Jim Cramer is (only half-jokingly) calling for a "half a trillion dollar buyback" to defend the stock (CNBC Fast Money). Watch: whether the $500B financing SPV clears antitrust review (the WSJ reported it's being examined); the January-quarter margin print.
  • AVGO (Broadcom). Bull: AI-semi revenue +221% YoY to $16.7B, guided to double in 2027 and again in 2028; Anthropic and OpenAI becoming top-2 XPU customers; keeps the higher-value Google inference socket; trades ~10x 2028 EPS on ~100% growth (Bloomberg Intelligence). Bear: lost Google's V8 training chip to MediaTek; margins compressing as XPUs scale; 2027 guide seen as light; fell ~6% and it was the "first major AI name to actually miss figures" (Schwab Network). Watch: the fiscal-year $100B AI-revenue target and whether the networking business (~40% of AI revenue, including Tomahawk optics) can offset accelerator margin dilution (Futurum Equities).
  • MRVL (Marvell). Bull: up 143% YTD, the "hot money" custom-silicon name, and (per CNBC) a beneficiary as hyperscalers move away from Nvidia to make their own chips (CNBC Fast Money). Bear: now caught in a three-way "ridiculous zero-sum game" with Broadcom and newly Nvidia-backed MediaTek; no fresh design-win disclosure surfaced this week. Watch: any Marvell reaction to the Nvidia and MediaTek tie-up encroaching on its ASIC-services turf.
  • GOOGL (Alphabet). Bull: the custom-silicon reference customer, now confident enough to split its TPU across three designers; TPU volumes ~9M units next year rising to 15–16M in 2028; also dodged a court-ordered break-up this week (The Information's TITV). Bear: Gemini's reputation "very much in the conversation right now, just not for the right reasons"; multi-sourcing adds execution risk (MediaTek must deliver V8). Watch: MediaTek's V8 ramp execution late this year; whether Gemini's next flagship reinserts Google into the frontier race.
  • MediaTek (2454 TT; not US-listed). Bull: $3.5B Nvidia investment plus NVLink Fusion access; AI-chip business ~$2B this year targeting ~15% of an $80B market next year; V8 training win at Google worth $14–16B next year (Bloomberg Tech Special Edition). Bear: unproven at data-center scale (a smartphone-SOC house moving up-stack); execution-dependent; supply-chain constrained ("not just HBM... substrate," per Tsai). Watch: first V8 volume shipments; whether the Nvidia partnership actually accelerates time-to-market. (Note: MediaTek trades in Taipei, outside this letter's US/UK/CN data coverage, and is covered here for context.)
  • ARM (Arm Holdings). Bull: CPUs "table stakes" in every AI factory (Vera, Graviton run on Arm); customer-base diversification as everyone builds custom chips is "only good"; edge AI a "sweet spot"; SoftBank as strategic anchor shareholder (No Priors). Bear: Haas himself warns of a 3 to 5 year supply-constrained environment and data-center-buildout bottlenecks that throttle the whole sector's growth. Watch: v9 royalty mix inside hyperscaler ASICs; whether faster AI-assisted chip design compresses the design cycle in Arm's favor.
  • AMZN (Amazon). Bull: Trainium is "sold out for years to come and improving with each generation," yet Amazon still committed to 2M more Nvidia GPUs, the both/and story (The Six Five). Bear: no fresh Trainium 3 or Project Rainier volume color this week. Watch: Trainium 3 ramp detail; any external-Trainium-sales disclosure.
  • Credo (CRDO), Astera Labs (ALAB). Bull: Credo delivered ~115% growth (~68% gross margins) and trades ~14.6x 2028 EPS; Astera remains one of the fastest-growing AI-infra chip names on rising content-per-rack (The AI Investor Podcast). Bear: both re-rating down hard with the "hardware jitters"; Astera hit ~100x earnings before pulling back to ~48x, and Credo sold off despite a beat (Futurum Equities). Watch: whether connectivity content-per-accelerator keeps scaling as XPUs grow.
  • MSFT (Maia), META (MTIA), Alchip. No dedicated podcast commentary surfaced on Microsoft's Maia, Meta's MTIA, or Alchip this week, as has been the case for several weeks. Flagging honestly rather than inventing.

Read-throughs

  • HBM memory is now the margin story, not just the toll booth. The most important read-through this week: Nvidia's own margin guide-down was blamed "mostly memory," and hyperscalers are increasingly buying HBM directly from the memory makers, routing that high-margin dollar around Nvidia (The Circuit). Watch SK Hynix, Micron, Samsung: they appear to have gained pricing leverage over the accelerator makers. Related: Nvidia's NVHBM (custom HBM on an Nvidia base die, now extended to NVLink Fusion partners) is a way to claw some of that value back.
  • The China memory wildcard. Dan Niles flagged an underappreciated threat: China's CXMT (DRAM) planning to raise wafer starts from 300k to 500k by end of next year, and YMTC (NAND) aiming to be bigger than Samsung and Hynix in NAND by end of next year. If they hit those goals, the "memory shortage through 2030" thesis that's propping up memory equities could break (Excess Returns).
  • TSMC packaging and the buildout bottleneck. Arm's Haas and The Circuit's hosts both put the binding constraint at wafers, CoWoS packaging, and, increasingly, the physical data-center buildout (labor, power, permits, local opposition), a 3 to 5 year governor on the whole complex. Calendar 2027 stays sold out; capacity meaningfully loosens only in 2028.
  • Connectivity content per accelerator (Astera, Credo, Broadcom and Marvell networking) remains the offset to accelerator margin dilution. Broadcom's networking is ~40% of AI revenue, and its move into silicon photonics (via its Grok acquisition) mirrors Nvidia's (The Morning Market Briefing).
  • The financing plumbing is the fault line under everything. This got heavy coverage and deserves its own callout:
    • David Bahnsen (The Bahnsen Group) walked through Nvidia's 10-Q: three customers = 54% of revenue (21/17/16%) and 64% of receivables; AI labs carry purchase-order commitments "over 10 times their total revenue." His thesis: this is now "a financial markets story rather than a technology story," and it "works until it doesn't" (The Dividend Cafe).
    • Dan Niles is watching hyperscaler credit default swaps every morning; even Nvidia's CDS now trades above the average North American investment-grade CDS. He expects a Fed rate hike on Sept 16 and is bearish into November (Excess Returns).
    • Aswath Damodaran and Alvin Wang Graylin (relaying the WSJ): ~$3 trillion of off-book AI debt, with data-center physical value only 5–10% of the collateral and 20 to 30 year securitizations against 4 to 5 year chip lives, "bigger than the 2008 subprime problems." Plus token-price deflation (frontier prices $50–60 down to $30–40; open-source now ~70% of OpenRouter traffic) squeezing lab revenue (Paradigm Shock).
    • The measured counter, from UBS CIO Americas Ulrike Hoffmann-Burchardi (Alpha Exchange): AI-capex rate elasticity is "very, very low", it'd take more than 50bps of rate moves to dent it, since hyperscalers already pay "three times the salary of an electrician" and "twice the base rate for electricity." Her key line for risk managers: circular financing "is not a risk as a catalyst. The catalyst will be if demand disappoints." And AI is now a hidden factor across seemingly-diversified portfolios (tech, utilities, REITs, EM semis, commodities all share it) (Alpha Exchange). (For the record: an automated tag initially credited this to Paul Tudor Jones. It's Hoffmann-Burchardi, who earlier in her career worked at Tudor.)

What changed vs last week

  1. The Google diversification thread got hard confirmation. Last week this was a rumor about Marvell and MediaTek being added as second and third TPU designers. This week Broadcom confirmed the specific split on its call, MediaTek = V8 training, Broadcom = inference, with dollar figures ($14–16B of MediaTek Google revenue next year). Rumor became design-win fact (The Information's TITV).
  2. A genuinely new strategic vector: Nvidia embracing XPUs. The $3.5B MediaTek investment is a new move, not a carry-forward. Jensen went from dismissing custom chips to building the on-ramp for them (and taxing the rest of the rack). This reframes the entire bull and bear debate.
  3. The margin question turned concrete. Last week's standing bear case was theoretical ("CUDA's moat diminished"). This week we got an actual number: Nvidia's own 75% to 72% gross-margin guide, blamed on memory. The bear case now has a data point.
  4. The financing plumbing remains the single most important dial on the board, now with harder numbers behind the worry.