NVDA News
32 R40 articles on NVIDIA Corporation (NVDA). Latest, Sep 30, 2026: CoreWeave's 3.8x Rubin Figure Is the Training Row, Not the 4.8x Inference One.
R40 Coverage
CoreWeave's 3.8x Rubin Figure Is the Training Row, Not the 4.8x Inference One
CoreWeave put Vera Rubin NVL72 into limited availability on 30 September. The 3.8x circulating is its training row; Cognition measured 4.8x for inference.
Apple Measured the 512GB Mac Studio's AI Claim on a 7GB Model
Apple's Mac Studio sells 512GB of unified memory. Every AI figure on its page is footnoted to a 14-billion-parameter 4-bit model, about 7GB of weights.
NVIDIA Qualified Three Batteries. Its Own Spec Says Sizing Is Irrelevant to Passing.
NVIDIA's DSX Ready qualified Tesla, Hitachi and LGES batteries on 21 September. Its own spec says battery sizing is irrelevant to passing the twelve tests.
SemiAnalysis Calls Its MI355X Chart Apples-to-Apples. Its Own Page Says It Is Not.
SemiAnalysis called its MI355X-vs-GB300 chart apples-to-apples. Its own page says disaggregated configs are not, and has GB300 2.3x to 5.4x cheaper per token.
Microsoft Already Told You It Builds a Gigawatt a Quarter. The 26-Gigawatt Leak Is Six More Years of It.
Microsoft discloses a gigawatt a quarter and $41bn of capex behind it, never a fleet total. The leaked 26 GW is that rate continued, and costs about $1.07tn.
NVIDIA Filed an 8-K to Buy Hugging Face for $11.9 Billion. It Filed Nothing to Pay Groq $17 Billion.
The DOJ is asking whether NVIDIA's Groq licence was an acquisition. NVIDIA's own filings book $17.0bn of consideration, $14.4bn of it as goodwill.
Huawei's Best AI Chip Is Indicated at 250,000 Yuan. The B200 Comparison Came From One Unnamed Source.
Two exclusives priced Huawei's Ascend 950DT on 10 September. The 250,000-yuan figure is one number reported twice, and only one of them mentions Nvidia's B200.
Nvidia's Up-To-2GW Australian Buildout by 2027 Names an 800MW Campus IREN Energises From 2028.
Two of Nvidia's eight Australian partners named megawatts. IREN's 800MW Bundey energises from 2028; CDC's 1,350MW is its own estate. The 2,150MW sum fails.
The Same 87.5% on ARC-AGI-1 Cost $4,560 in 2024 and 30 Cents in 2026. ARC-AGI-4 Lands in 2027.
A deep dive on ARC-AGI: three versions, seven years of scores, a like-for-like 15,200x cost collapse, and how much of it is hardware rather than software.
Anthropic Doesn't Trade. It Is Already a $53.4B Line in Amazon's Quarter and Two-Thirds of Microsoft's Beat.
Amazon booked $53.4 billion primarily from its Anthropic stake, and a $3.2 billion gain was two-thirds of Microsoft's beat. The IPO reprices both at once.
Anthropic Is Helping Google Build a $67.9B Answer to Nvidia's $108.5B Financing Machine.
Anthropic's million-TPU commitment is helping Google turn an internal accelerator into a financed merchant platform. Nvidia's guarantee book remains larger.
Meta Is Building Toward 14 GW. Its Rentable Compute Is Still a Thin Slice.
Meta targets 14 GW by 2027 while remaining a major compute buyer. This deep dive maps its Titans, silicon, workloads, utilization and rentable capacity.
Nvidia's Megawatt Got 60% More Expensive and 30 Times More Productive in One Generation.
Nvidia disclosed both curves: a megawatt went $25m to $40m and got 30x more productive. Revenue per megawatt holds only if token prices fall less than 30x.
Nvidia Put $158.5 Billion Behind the AI Labs. It Expects to Sell Them $167 Billion Next Year.
A quarter of Nvidia's next year leans on its balance sheet — roughly $167bn against the 70% guide, behind which sits $158.5bn of its own equity and guarantees.
Nvidia Guided a Year Ahead for the First Time — and Cut Its Margin Floor to 71%
Nvidia doubled revenue to $96.2B, then did two things it never has: guided a full year it has not started, and told the market gross margin bottoms at 71-72%.
Anthropic Has Contracted 1.19 GW. Half of It Costs $2.37M per MW-Year, the Other Half $15.46M.
Four reported Anthropic capacity deals total 1,185 MW for $83.1B, and split by term they are two products: $2.37M per MW-year on 20-year leases, $15.46M on six.
Nvidia Returned a Record $26 Billion. At 44 Times Free Cash Flow, a Full Payout Would Still Yield 2.3%.
Nvidia returned a record $26bn last quarter, 120% of free cash flow. At 44x free cash flow that is a 1.8% yield, and a 100% payout would reach only 2.3%.
A Gigawatt of Data Centre Costs $60 Billion. Nvidia Says $40 Billion of It Is Nvidia.
Jensen Huang priced both sides on the call: a gigawatt of data centre costs $60bn, and $40bn of that is Nvidia. Five years ago it was $30bn and $3–5bn.
OpenAI’s Jalapeño Beat GB300 on Its Test. NVIDIA’s $63 Bear Case Needs Three More Things.
OpenAI's Broadcom-built Jalapeño beat GB300 on throughput per watt. Our $63.15 NVIDIA bear case still needs capex digestion, share loss and a de-rating.
Cisco Added Supermicro to Its AI Factory.
Cisco adds Supermicro's Vera Rubin systems to its Secure AI Factory in October — with no customer order, rack count, contract value or margin disclosed.
Waymo Is Driving on 5nm. Xiaomi Has Validated 3nm. Tesla Is Moving From 7nm AI4 to a Reported 2nm-Class AI5.
The robotaxi silicon ladder runs from Tesla's 7nm AI4 to a 2nm-class AI5, yet Waymo's mature 5nm ASIC leads the production test: it carries passengers today.
The Non-Linear Rise of AI Tokens: Why Cheaper Intelligence Creates More Demand
Token prices are collapsing while usage explodes. The cause is a sequence of demand regimes — chat, coding agents, persistent multi-agent systems — each bigger.
NVIDIA Q2 FY2027 Earnings Preview
Nvidia reports Wednesday with consensus 1.3% above its own $91.0B midpoint. It has beaten four straight times by 5-10% and the stock fell after every one.
Waymo's Custom Chip Cuts Cost, but a $125,000 Car Is Still No Threat to Tesla's Scale.
Waymo cut $75,000 from its robotaxi and its 5nm ASIC is real engineering — but it is a sensor pre-processor on a $125,000 car, four to six times a Cybercab.
Cerebras Shipped a 132GB Wafer-Scale Inference Machine Without HBM. Its Models Need 50,000.
Cerebras' CS-4 proves inference silicon can skip HBM: 750 PFLOPS and no high-bandwidth memory. But its 132 GB of SRAM faces models needing 25,000-100,000 GB.
Google's $12.2B Marvell Warrant Vests Against $120B of Revenue—Not Equity Value.
Marvell's Google warrant covers 58.97M shares, but only 1.36M vest on time. The rest need one $500M revenue tranche each — $120B total, against an $11.5B guide.
China Admitted 20,000 H200s. NVIDIA's Outlook Assumes Zero Revenue From Them.
The FT reports ByteDance and Tencent each received about 10,000 H200s — $0.5-0.8B at any plausible price, inside the ±2% tolerance on Nvidia's own $91.0B guide.
Nvidia's Disclosed Stock Portfolio Grew $45 Billion Last Quarter. It Bought One Thing.
Nvidia's Q2 13F shows eight positions worth $63.4B, up $45B in a quarter. Seven share counts are unchanged: the move is one new position and Intel tripling.
In February We Warned That Beating Wouldn't Be Enough. Nvidia Beat, Then Grew Another 20%.
Nvidia beat the consensus our February preview named by 3.9% on revenue and 15.8% on EPS, then grew 20% again the quarter after. Half that call we can grade.
The AI Capex Bet, Six Months In: What the Mag 7 Actually Spent (and What the Market Learned)
Our February call was a $700B AI capex arms race. Six months and a full Q2 earnings season later, actual guidance has blown past that number, free cash flow is cratering at multiple hyperscalers, and the market has started rewarding capex very differently depending on who's spending it. Here's what changed.
Nvidia Reports Next Week. Wall Street Has Never Been This Confident.
Nvidia's Q4 FY2026 earnings drop February 25. Analysts expect $65.6B in revenue and $1.52 EPS — but the stock is flat in 2026. Here's what the numbers say and what could go wrong.
The $700B AI Bet: Inside the Mag 7's Record 2026 CapEx Arms Race
The Magnificent Seven are set to spend $700B+ on AI infrastructure in 2026, up 60% from 2025. We break down who's spending what, the supply chain winners, and why the market is spooked.
February 2026
Jensen Huang: Demand explosion easing GPU depreciation fears
Old GPUs increasing in price; hyperscaler capex acceleration benefits NVDA and neoclouds.
NVDA CEO on compute demand
OpenAI/Anthropic revenue could 4x with more compute; massive constraint driving spend.
Nvidia shares surge 7-8% leading tech rebound
Big Tech's $650B AI capex plans (incl. Amazon's $200B) drive gains; Jensen Huang highlights compute constraints and rising GPU values.
NVDA Short-Term Pressure
Reports of AI customers exploring alternatives and China H200 delays; dominates AI chips, bullish ratings ahead of earnings.
NVDA Rejection from Supply
Sharp rejection from 480–490; bounce from demand but rallies sellable; weak sell bias unless reclaim mid-450s.
January 2026
NVDA Not Finished
Structural selling; growth concerns and valuation compression continue downside.