Lead Analysis
A Hyderabad Startup's 9,000-GPU Order Shows Where Nvidia Demand Is Spreading Next
AM Intelligence, a Hyderabad-based AI-infrastructure company backed by renewable-energy producer Greenko Group, placed a binding order for 9,000 Nvidia Vera Rubin NVL72 rack-scale systems, with delivery slated for the first quarter of 2027, Bloomberg reported Tuesday. The order sits inside the company's broader $8 billion plan to build 1 GW of computing capacity at a Hyderabad facility aimed at training trillion-parameter models, with initial capacity already sold to an undisclosed US customer under a non-disclosure agreement.
"Customers are hunting desperately for capacity … In global token economics, energy prices play a huge part. We're one of the lowest-cost AI computation infrastructure players globally." — Mahesh Kolli, Founder & President, Greenko Group, via BloombergKolli's pitch rests on Greenko's access to low-cost renewable power, letting AM Intelligence undercut Western and Gulf rivals on compute pricing even as it competes for the same GPU allocations feeding hyperscaler build-outs. The company plans to fund its buildout with a mix of debt and equity, citing prior experience tapping bond markets, and is targeting 5GW of capacity by 2030 across India and Europe. The order lands a day before Nvidia's own fiscal second-quarter results, when Wall Street will be parsing how quickly the Vera Rubin ramp is translating into bookings beyond the usual hyperscaler customer base.
Markets
Daily Top-10 Movers
Prices are Tuesday, August 25, 2026 closes at 4:00 PM EDT. Ranked by absolute price change among semiconductor stocks with market cap ≥$5B.
★ highlighted = today's news-driven movers.
Companies & Financials
Navitas Doubles Down on AI Power as Semtech's Data-Center Bet Pays Off
- Navitas Semiconductor agreed to acquire power-delivery startup Claros for up to $232.8 million, roughly $216 million of it payable at closing in cash and stock, per GlobeNewswire. Why it matters: the deal extends Navitas's power-delivery reach from the grid to the processor itself, and the company estimates it more than doubles its addressable market to over $8 billion, positioning a sub-$3 billion power-semiconductor name to capture a bigger share of AI data-center buildouts than its size would suggest.
- Semtech reported record fiscal second-quarter revenue of $341.9 million, up 17% sequentially, with adjusted EPS of 71 cents versus roughly 61 cents guided, and issued third-quarter revenue guidance of about $410 million, some $50 million above consensus, per Yahoo Finance. Why it matters: data-center revenue hit a record $100 million, up 91% year-over-year, on 800G and early 1.6T ramp, showing that optical and copper interconnect suppliers are capturing AI infrastructure spending well outside the GPU and memory names that dominate the headlines.
Memory Desk
Gartner Sees Memory Revenue Nearly Quadrupling as Mizuho Flags a De-Specing Risk
- Gartner forecasts worldwide semiconductor revenue will jump 92% to about $1.6 trillion in 2026 from $809 billion in 2025, with memory revenue surging from $220.1 billion to $837.3 billion over the same span, per Benzinga, citing director analyst Ben Lee. Why it matters: Gartner expects DRAM revenue to climb 246.6% and NAND revenue 371.9% this year, meaning memory alone could account for 54% of total chip revenue in 2026, up from 27% last year — a concentration that raises the stakes for every memory-cycle wobble.
- Mizuho's Vijay Rakesh trimmed his price targets on Micron to $1,300 from $1,375 and on SanDisk to $1,875 from $1,900, while keeping Buy ratings on both, per TipRanks. Why it matters: Rakesh's cut followed a meeting with SemiAnalysis's Doug O'Laughlin flagging that GPU and ASIC makers may cap future HBM content at 4-8Hi rather than 12Hi, a de-specing risk that would compress the memory-content-per-chip story Wall Street has been pricing into Micron and SanDisk shares.
Supply Chain, Equipment & Fabs
Memory Is Eating Cloud Capex as Nvidia's Rack Redesign Chases Faster Rubin Deployment
- DRAM and NAND combined will account for 47% of major cloud providers' capital spending in 2026, rising to 68% in 2027, as total CSP capex surges 98% year-over-year this year and a further 50% next, per TrendForce, via Evertiq. Why it matters: server DRAM contract prices are projected to rise roughly 270% cumulatively in 2026 and HBM contract prices 70-140% in 2027, meaning hyperscalers are increasingly spending their AI budgets on memory rather than new compute silicon or equipment orders.
- Jefferies projects Vera Rubin will contribute about 12% of Nvidia's GPU revenue in fiscal third-quarter 2027, rising above 40% by the fourth quarter, while a redesigned rack architecture is expected to cut compute-tray assembly time from roughly two hours to five minutes, per TradingKey. Why it matters: faster physical assembly directly speeds how quickly data-center operators can convert Nvidia shipments into billable capacity, a deployment-efficiency lever that matters as much to the AI buildout's pace as wafer supply itself.
Policy & Geopolitics
Washington's Chip-Policy Signals Get Murkier as SEC Digs Into an AI Fund's Collapse
- The SEC sent subpoenas to Goldman Sachs, JPMorgan, Citigroup and Bank of America seeking records on their financing and trading with Situational Awareness, the AI-focused hedge fund that lost about $35 billion of its roughly $45 billion peak assets after using up to 400% leverage, per CNBC. Why it matters: the fund's forced unwind in July has been cited as a contributor to recent volatility in AI-linked equities, and the probe could surface how much leveraged exposure major prime brokers still carry to concentrated AI bets.
- The Trump administration is reportedly preparing to let Apple source memory chips from China's CXMT and YMTC after a planned September meeting between President Trump and China's Xi Jinping, reversing Commerce Secretary Howard Lutnick's earlier public opposition, per WCCTech. Why it matters: the unconfirmed report is what drove Monday's roughly 7% slide in Micron and SanDisk shares, and a confirmed reversal would mark a real shift in how far Washington is willing to bend its memory-reshoring stance for diplomatic leverage.
Analyst Corner
SemiAnalysis Weighs OpenAI's Chip Against Blackwell as Equipment Demand Resets
- SemiAnalysis (published 21 hours ago): Bryan Shan, Myron Xie, Jordan Nanos and colleagues argued that OpenAI's self-designed Jalapeño ASIC beats Nvidia's Blackwell on select total-cost-of-ownership and throughput-per-megawatt metrics, a direct comparison against Rubin that bears on how seriously hyperscaler custom silicon now threatens merchant GPU economics.
- Morgan Stanley (background reading, published August 21): the firm's semiconductor-equipment team raised its 2027 wafer-fab-equipment outlook, citing stronger-than-expected DRAM investment signals from Advantest and Tokyo Electron's earnings season and naming Advantest its top equipment pick — a read on capital spending that still frames how this week's memory-capex data should be interpreted.
Calendar / Week Ahead
Nvidia's Earnings Wednesday Set the Tone Before Marvell Reports Thursday
- Wednesday, Aug. 26 — Nvidia reports fiscal second-quarter results after the closing bell, with Wall Street consensus near $92 billion in revenue against the company's own $91 billion guidance, per TradingKey.
- Thursday, Aug. 27 — Marvell reports fiscal second-quarter results, the first test of its AI-networking momentum since Susquehanna's Tuesday price-target increase, per Seeking Alpha.
|
