Marvell agrees $115 billion Lumentum deal as chip stocks slide on AI spending worries
Anthropic's AI infrastructure commitments climbed to nearly $900 billion, Amazon disclosed an $80 billion loan facility and Oracle priced $10 billion of bonds. Onsemi agreed to buy Ferric for $600 million, yet its shares fell 6.6%, and the Kospi slid 3.36% at the open on concerns over AI infrastructure spending.
News +0.54Market no trading session in this window: the score is the news reading
Key notes
115bn USD
Marvell's price for Lumentum
Marvell agreed to the largest AI-infrastructure chip deal of the window. It will buy Lumentum for $115 billion, a deal Reuters called costly. Business Insider reported that Marvell's stock dipped on the news.
Chip and AI stocks sold off across markets. BigGo Finance reported the Philadelphia Semiconductor Index fell 3.37% and TSMC's ADR dropped below $400 as investors grew wary ahead of the midterms. Korea JoongAng Daily reported the Kospi opened 3.36% lower on concerns over AI infrastructure spending.
Anthropic's infrastructure commitments grew. Mezha.net reported that the commitments climbed to nearly $900 billion, and Indian Television Dot Com reported that the company posted third-quarter revenue alongside them.
Hyperscalers raised more debt for the buildout. Cloud Computing News reported that Amazon disclosed an $80 billion loan facility, and International Financing Review reported that Oracle priced $10 billion of bonds amid a surge in hyperscaler supply.
Onsemi bought into AI power while its stock fell. TradingView reported that it will acquire Ferric for $600 million in cash and raised its 2030 AI data center revenue target. GuruFocus reported that the shares fell 6.6% after the Q3 results and the acquisition.
Grid build-out for AI moved forward in the UK. Data Center Dynamics reported that National Grid bought land in Teesside for a 3.2GW substation to power an AI data center. Separately, Unite.AI reported that Envision deployed 800V DC power based on solid-state transformers at its Ulanqab AI data center.
Data centers · 33 storiesSites, colocation, cooling, capacity leases.Samsung SDS breaks ground in Gumi; Bell's Saskatchewan site faces First Nations scrutinyNew sites appeared in Korea and Vietnam, while Bell's AI data centre drew criticism over permitting and political ties. Foreign Policy Journal reported an 800 MW financing gap at Applied Digital.Frontier AI sites: power against capital costEach point is a site in Epoch AI's frontier register with both a stated power capacity and a stated capital cost — the further above the trend, the more each megawatt cost to build.THE ELEC reported Samsung SDS began construction of its Gumi AI data center and plans to expand it to 120MW.THE ELEC, Korea Electronics Industry MediaCBC reported First Nations are demanding Bell release an environmental assessment of its new AI data centre.CBCW.Media reported a Vietnamese province is considering a US$700 million AI data center project.W.MediaOpen 33 stories →
Hyperscalers · 15 storiesCapex, power procurement, model-capacity deals.Amazon, Oracle and Meta tap debt markets for AI buildoutAmazon disclosed a loan facility and Oracle priced bonds, while Oracle shares fell. Morgan Stanley's chief executive put AI infrastructure spending at $15 trillion over ten years.Hyperscalers, indexed over the last 10 sessionsClosing prices on each company's home exchange, indexed to 100 at the first session shown, for the 5 hyperscalers that moved most over the window.Cloud Computing News reported Amazon disclosed an $80 billion loan facility as AI infrastructure spending rises.Cloud Computing NewsInternational Financing Review reported Oracle priced US$10 billion of bonds, and NDTV Profit reported Oracle shares fell 5%.International Financing ReviewNDTV ProfitBigGo Finance reported Morgan Stanley CEO Ted Pick expects AI infrastructure to absorb $15 trillion over ten years.BigGo FinanceOpen 15 stories →
Semiconductors · 43 storiesAccelerators, memory, foundry, packaging.Marvell's $115 billion Lumentum deal leads a day of chip M&AOnsemi added Ferric and said its AI data center revenue would double again next year. South Korea plans to double AI and chip spending in 2027.Chipmakers this window, by move0 of 8 tracked chipmakers rose over the window; the change is close to close on each company's home exchange, in its own currency.Reuters reported Marvell will buy Lumentum for $115 billion.ReutersBigGo Finance reported ON Semiconductor expects its AI data center revenue to double again next year, and TradingView reported it will buy Ferric for $600 million.BigGo FinanceTradingViewTaipei Times reported South Korea will double its spending on AI and chips next year.Taipei TimesOpen 43 stories →
Semi equipment · 4 storiesLithography, deposition, etch, test, export rules.Nitrogen chiller and AI-based defect inspection target chipmakingA zero-GWP chiller and AI image enhancement for flip-chip inspection were reported, and Edwards' New York pump plant drew more coverage.Toolmakers from window open to close0 of 5 tracked toolmakers closed the window above where they opened it; each line is one company's close, indexed to 100 at the open.gasworld reported a nitrogen-based chiller targeting the semiconductor market with zero-GWP cooling.gasworldSemiconductor Engineering reported AI image enhancement improves detection of flip-chip interconnect defects.Semiconductor EngineeringConstruction Review reported Edwards Vacuum opened its $319M New York factory to produce semiconductor pumps.Construction ReviewOpen 4 stories →
New research
Speculative decoding papers report the largest inference speedups
SpecPilot (UC Berkeley, Alibaba) reports a 10.9× speedup over autoregressive decoding. FactorSpec reports 10.5× on tokens per verify step, and Alibaba's SpecOPD reports 3.6× on 8B–125B targets. Only Duke and Meta's BasisKV, with a 1.4× end-to-end speedup, released code, while Stanford and Google DeepMind's Hinted Scaling Laws targets forecasting from sub-frontier runs.
Which fields movedpapers today ● vs 30-edition average ○
01345.6 papers
How open was it
Weights released0 / 8
Code released1 / 8
Who published2 lab · 3 mixed · 3 academic
Reported gains today, as multiples
5 of 8 kept papers report their headline result as a speed-up multiple; papers reporting in other units are not put on this axis.
The papers · headline result, scale, what was released
AI + Energy & Sustainability
Frontier AI power on the record, by country8 rows · 1 source
Stated power capacity summed across the frontier sites Epoch AI records in each country — where the grid load of the build-out actually sits.Epoch AI · Frontier Data Centers Hub
2 charts built from 5 stated figures across 15 stories, the first planned by the composer from the record. Also had the data for what today covered · topic by region, when the energy stories landed, pressure and relief by topic. Held back: power committed in deals that state a figure (no deal in this edition states the capacity it commits); power disclosed per edition · last 7 (only 1 of the last 7 editions put a number on it).