SemiconductorsFEATURE

Google's $120 Billion Marvell Deal, Explained: How Custom AI Chips Are Eating Into NVIDIA's Lead

E
EffectStory 編輯部Editorial Team
Published · Updated
The widely reported '$120 billion' Google-Marvell deal is not an order or a revenue figure — it is a cumulative custom-chip purchase threshold Marvell must hit by fiscal 2033 to fully vest warrants Google holds. Marvell's own guidance that this Google revenue won't ramp meaningfully until fiscal 2029 triggered an 8%-plus stock drop, even as Amazon, Meta, and Google all scale in-house AI silicon against an NVIDIA that still commands roughly 70% of the AI chip market.

Is the "$120 Billion" an Order or a Performance Target? The Real Structure of the Google-Marvell Deal

The $120 billion figure is neither an order nor annual revenue — it is the cumulative custom-chip purchase threshold Marvell must reach by fiscal 2033 for Google's warrants to fully vestCITE:E1. According to an analysis published by Next Platform, the number comes from a vesting schedule split into 240 tranches: each time Marvell recognizes $500 million in custom-chip revenue from Google, one tranche unlocks, and 240 tranches multiplied by $500 million equals $120 billion — a "hit the target, get the equity" structure rather than a purchase commitmentCITE:E1.

The mechanism also ties the two companies' financial interests together. Marvell has granted Google warrants to purchase up to 58.97 million shares at a strike price of $206.58 per share, a stake worth roughly $12.2 billion if exercised in fullCITE:E2. Should the warrants fully vest, Google would become approximately Marvell's fifth-largest shareholder — turning a chip-supply relationship into a direct equity stake between supplier and customerCITE:E2.

Why Did the Stock Fall After a Blockbuster Deal? The Gap Between Expectations and Timing

Marvell shares fell more than 8% following its August 28, 2026 earnings report, despite the underlying results being strongCITE:E3. The drop wasn't about the quarter itself — CEO Matt Murphy said that Google-related revenue is not expected to ramp meaningfully until fiscal 2029, later than the market had anticipated, and the company gave no new fiscal 2029 financial forecastCITE:E3. Investors read that combination — a long vesting runway to fiscal 2033 paired with no meaningful revenue until fiscal 2029 — as confirmation that the $120 billion headline describes a multi-year trajectory, not near-term cash flow.

Has Google's TPU v7 "Ironwood" Caught Up to NVIDIA?

Google's TPU v7, codenamed "Ironwood," is now live and delivers specifications that roughly match NVIDIA's flagship AI acceleratorsCITE:E4. Each chip delivers 4.6 PFLOPS of compute paired with 192GB of HBM, and Google assembles 9,216 of these chips into a single superpod rated at 42.5 exaFLOPS, aimed at large-scale training and inference workloadsCITE:E4.

Amazon Trainium, Meta MTIA: How Are Cloud Giants Scaling In-House Silicon?

Amazon has already deployed more than 1 million of its self-designed Trainium2 chips through Project Rainier, while Meta is co-developing its MTIA chip with Broadcom on a 2-nanometer processCITE:E5CITE:E6. Project Rainier powers training and serving for Anthropic's Claude, and Anthropic has committed to more than $100 billion in compute spending over 10 years, with Amazon reserving up to 5GW of Trainium capacity for that relationshipCITE:E5. Meta's first-phase MTIA deployment with Broadcom exceeds 1GWCITE:E6. Broadcom, the custom-ASIC design partner behind that effort, reported AI semiconductor revenue of $10.8 billion in its fiscal 2026 second quarter with a backlog exceeding $30 billion — a scale larger than Marvell'sCITE:E6.

How Severe Is the ASIC "Encroachment"? Is NVIDIA's Threat Immediate?

NVIDIA still holds roughly 70% of the AI chip market, but custom ASICs are growing far faster than commodity GPUsCITE:E7. An estimate reported by Tech Times puts 2026 custom-ASIC shipment growth at approximately 44.6% year-over-year, compared with about 16.1% for commodity GPUs — a gap driven by cloud providers' pursuit of cost control, supply autonomy, and workload-specific optimization at scaleCITE:E7.

Custom Silicon at a Glance

PlayerChip / ProgramScale SignalKey Figures
GoogleTPU v7 "Ironwood"9,216-chip superpod4.6 PFLOPS/chip, 192GB HBM, 42.5 exaFLOPSCITE:E4
Google–MarvellWarrant-linked supply dealVesting through FY2033240 tranches × $500M = $120B; 58.97M shares @ $206.58 (~$12.2B)CITE:E1CITE:E2
AmazonTrainium2 (Project Rainier)>1 million chips deployedAnthropic: 10-yr, >$100B commitment, up to 5GW reservedCITE:E5
MetaMTIA (with Broadcom, 2nm)First-phase deployment>1GWCITE:E6

Market Share and Growth

MetricValueSource
NVIDIA AI chip market share~70%CITE:E7
2026E custom ASIC shipment growth~44.6% YoYCITE:E7
2026E commodity GPU shipment growth~16.1% YoYCITE:E7
Broadcom AI semiconductor revenue (FQ2 FY2026)$10.8BCITE:E6
Broadcom AI semiconductor backlog>$30BCITE:E6
Marvell stock move after Aug 28, 2026 earnings-8%+CITE:E3

Taken together, the evidence points to a structural rather than sudden shift. The Google-Marvell warrant mechanism binds supplier and customer through fiscal 2033, but Marvell's own guidance shows the associated revenue stays muted until fiscal 2029 — which explains the market's negative reaction despite the deal's headline sizeCITE:E1CITE:E3. Meanwhile, Amazon and Meta are already operating at scale — over 1 million Trainium2 chips and a 1GW-plus MTIA deployment, respectively — while Google's TPU v7 Ironwood posts specifications comparable to NVIDIA's own acceleratorsCITE:E4CITE:E5CITE:E6. NVIDIA's roughly 70% market share remains intact for now, but the reported gap between 44.6% ASIC growth and 16.1% GPU growth in 2026 shows custom silicon expanding from a smaller base at a faster rateCITE:E7.

📊 Evidence

FAQ

Is the "$120 Billion" an Order or a Performance Target? The Real Structure of the Google-Marvell Deal

The $120 billion figure is neither an order nor annual revenue — it is the cumulative custom-chip purchase threshold Marvell must reach by fiscal 2033 for Googl…

Why Did the Stock Fall After a Blockbuster Deal? The Gap Between Expectations and Timing

Marvell shares fell more than 8% following its August 28, 2026 earnings report, despite the underlying results being strongCITE:E3.

Has Google's TPU v7 "Ironwood" Caught Up to NVIDIA?

Google's TPU v7, codenamed "Ironwood," is now live and delivers specifications that roughly match NVIDIA's flagship AI acceleratorsCITE:E4.

Amazon Trainium, Meta MTIA: How Are Cloud Giants Scaling In-House Silicon?

Amazon has already deployed more than 1 million of its self-designed Trainium2 chips through Project Rainier, while Meta is co-developing its MTIA chip with Bro…

📎 Sources

  1. nextplatform.com
  2. investing.com
  3. bnnbloomberg.ca
  4. servethehome.com
  5. aboutamazon.com
  6. about.fb.com
  7. techtimes.com

Related data

Author's TakeEffectStory 編輯部

The Google-Marvell warrant structure is worth watching precisely because it converts a supplier relationship into an equity stake — Marvell only earns Google's backing as a shareholder if it actually delivers $500 million increments of custom-chip revenue, 240 times over, through fiscal 2033. That mechanism explains why Marvell's stock fell more than 8% right after the deal's scale became public: investors read Matt Murphy's own guidance — no meaningful Google revenue until fiscal 2029 — as confirmation that the $120 billion figure describes a multi-year trajectory, not near-term cash. Set against Amazon's already-operational fleet of over 1 million Trainium2 chips and Meta's 1GW-plus first-phase MTIA deployment with Broadcom, Google's TPU v7 Ironwood and its Marvell tranche structure look less like a lead than an attempt to catch up to a race already underway. The number to track next is whether Marvell's tranche count actually advances toward fiscal 2029, or stalls.

E
EffectStory 編輯部Editorial Team

Related

FEATURE

Space-Based AI Data Centers: What SpaceX, Google, and China Have Actually Launched So Far

Only two orbital AI computing efforts have hardware actually flying: Starcloud's single NVIDIA H100 satellite, launched in November 2025, and China's Three-Body Computing Constellation, which put 12 satellites into orbit on May 14, 2025, carrying 5 POPS of compute and 30 terabytes of storage. SpaceX/xAI's Starmind and Google's Project Suncatcher remain unlaunched, targeting 2027 at the earliest, with no compute figures yet disclosed.

EffectStory 編輯部 ·
FEATURE

AI Inference Prices Are Falling About 10x a Year — What the Token-Cost Data Actually Shows

Equivalent-capability AI inference prices have fallen roughly 10x per year across multiple benchmarks, with OpenAI's GPT-4o mini priced about 99% below its 2022 predecessor and DeepSeek-R1 undercutting OpenAI's o1 by roughly 27x. Yet Google now processes about 50 times more tokens monthly than a year earlier, and Gartner projects 2025 global generative-AI spending at $644 billion, up 76.4% — proof that cheaper tokens are fueling more total usage and spend, not less.

Nathan ·