New Ask Claude or ChatGPT about any stock, using Bullpen. Bullpen in Claude or ChatGPT. 7 days free
Bullpen.News From Edwin Dorsey
Get Daily Recap Feedback

NASDAQ

Nvidia CORP

NVDA

251 stories

$4.6T Market cap · 2026-11-06

Get NVDA news in Claude, ChatGPT or your own code. See plans

@aleabitoreddit 9 clicks

Serenity on X

$NVDA Jensen Huang: Surplus of memory supply is unlikely by 2029. The CNBC question referenced a Korean newspaper opinion on $KOSPI / Samsung / SK Hynix that predicted "new supply will leave the memory sector a red ocean in 2029." "One of the things that's really great about memory is that you can always use more of it", Jensen said, citing LPDDR usage in DCs. And responded "I don't think so" in terms of possibilities of oversupply in 2029. Just my personal opinion, Nvidia probably has the most visibility in the entire industry. Especially around supply procurement and emerging capacity, given they're the ones leading all the architectural shifts. On durability, you have Samsung already allocated majority of capacity into 2031 from SED and hyperscalers like $GOOGL are already reportedly securing memory LTAs into 2035 from KeyBanc. Yet there's endless "it's all over" sentiment on memory now. I'd probably trust Jensen's opinion here.
@convequity 5 clicks

Convequity on X

We think Huawei's control of lasers, optics and packaging could weaken pricing power at NVIDIA ($NVDA) and Broadcom ($AVGO). Our view: designing these components together lets Huawei choose embedded lasers based on what works best for the whole system, with less need to accommodate separate suppliers' products and margins. That makes lower-power embedded lasers easier to use and reduces dependence on scarce high-power external lasers. Huawei's 7.2 Tb/s Hi-ONE optical engine, which we highlighted in our recent report, shows the progress in connectivity behind this challenge. Better connections between processors can help an AI system use its computing capacity more efficiently, giving Huawei a way to compete beyond the performance of individual chips. If this lowers costs and deployment delays, customers in markets open to Huawei get a stronger alternative to systems built around Nvidia accelerators or Broadcom's custom chips and networking. At Convequity we think that could weaken both companies' ability to command premium prices.
@convequity 8 clicks

Convequity on X

Nvidia's ($NVDA) own new programming language makes it easier to move AI code to rival chips. In December 2025 Nvidia launched CUDA Tile. In a tile language the programmer describes work as blocks of data and the compiler works out how to run it on whichever GPU it is given. Our read: Nvidia no longer writes code around the CUDA core. Code written for one GPU generation increasingly breaks on the next, so a higher-level language was the only practical way forward. Tile code moves easily to other AI chips. On 30 September DeepSeek released an Ascend version of TileLang, the open tile language it uses, so every TileLang operation in its training now runs efficiently on Huawei's chips, not just Nvidia's. Few investors have noticed. Our view: once the code that matters is written in tiles, the switching cost that supports Nvidia's pricing power falls.
@aleabitoreddit 3 clicks

Serenity on X

$NVDA today announces capacity agreements with substrate, packaging, testing and optical engine suppliers. - Substrate agreements with Ibiden, Samsung Electro-Mechanics and Unimicron to expand by 2.5x - $AMKR to increase advanced packaging capacity by more than 300% by the end of 2028, with a new $1.3B agreement - Multibillion dollar capacity reservation advance to support Ibiden's plant in Japan - "By 2028, production allocated to $NVDA is projected to account for a significant portion of Samsung Electro-Mechanics' total annual package substrate production capacity" - KYEC with $1.1B capacity reservation in advance - SPIL for optical engine capacity for CPO by 15x by H2 2027 and double that for 2028 Well, that's a lot of news. As Jensen Huang puts it: "Demand is extraordinary. Capacity is the constraint."
@convequity 5 clicks

Convequity on X

AI models are now writing the kernels that made CUDA so hard to leave, and we think this breaks another part of $NVDA's moat. Kernels are the code that tells a chip how to carry out each basic step an AI model repeats, and how well that code is written decides how fast the model runs. Nvidia prebuilt a huge range of these through CUDA. Anyone building their own chip had to write that software from scratch, which was slow, hard and a big reason to think twice before starting a custom-chip project. OpenAI showed at Hot Chips in August that this is changing. On Jalapeño, its Broadcom ($AVGO) built chip, an internal version of Codex wrote attention and mixture-of-experts kernels that ran 1.5 to 1.8x faster than ones written by human experts. Within about three months of getting its first working chips, OpenAI had three open-weight models running on Jalapeño, even though none of them was designed for it. Our view: if AI can write the kernels, the software barrier that held back custom chips gets much lower, and we expect more companies to start their own chip projects. Investors still do not pay enough attention to that consequence. Deeper dive available at Convequity.
@aleabitoreddit 10 clicks

Serenity on X

Yeah not sure how anyone thinks it's over on AI semis like $NVDA GPUs, memory, to photonics, after $SPCX earnings. - Previous Q4 capex outlook ~$18.4B... Now $35–40 billion Q4 capex guidance, basically doubling previous outlook "Q3 and Q4 capex should be similar to Q2's $18.37B" from CFO in August. -Elon: "If everything goes right, we’ll have 20 gigawatts of incremental power next year" "High confidence of 10 GW next year, and 20 is difficult but not impossible - "I’m being as precise as possible about this. It is absolutely insane to be deploying and providing the power and all of the electrical equipment, the cooling, the batteries, the networking, and everything else that’s needed to turn on 220,000 GB300s a month" You're gonna need a lot of HBM, power, networking/optical components to make all of this happen. - Elon projects SpaceX could exit 2027 with an ARR of "$500 billion and $1 trillion by the end of 2027 When you see $SPCX double capex outlook next quarter, target another 10GW deployment by next year, and all the procurement needed to turn on 220,000 GB300's a month... Hard not to be bullish on AI semis.
@StableBread 6 clicks

Fajasy on X

I went through every Boost Run $BRUN contract with disclosed terms to see how much of its $2.6B+ of total contract value (TCV) they cover. Cohere, the Canadian developer of AI models for businesses and governments, agreed on September 30, 2026, to pay $BRUN $525.6M over the agreement’s term for $NVDA’s GB300 NVL72 racks, each with 72 Blackwell Ultra GPUs. Part of the $525.6M is a prepayment $BRUN hasn’t disclosed. But if we follow the 22% average $BRUN says its prepayments run, the prepayment would be ~$116M. Each rack’s term runs 60 months, and the first racks should start in Q2 2027. Cohere can terminate and get back everything it prepaid if it hasn’t accepted a minimum number of racks by July 15, 2027. $BRUN has disclosed terms for seven contracts, and the two largest are with Cohere and Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati: → Fluidstack (November 6, 2025): $63.6M over 24 months, or $31.8M/year. → 1,280 B300s, Charlotte (March 15, 2026): $116.1M over 36 months, or $38.7M/year. → 256 H200s, Raleigh (March 16, 2026): $3.7M over 12 months, or $3.7M/year. → Thinking Machines Lab (May 21, 2026): $471.7M over 36 months, or $157.2M/year. → June order, 240 GPU servers (June 8, 2026): $207.6M over 36 months, or $69.2M/year. → July order, 1,536 B300s (July 9, 2026): $222.5M over 48 months, or $55.6M/year. → Cohere (September 30, 2026): $525.6M over 60 months, or $105.1M/year. --- → Total: $1.61B, or $461.3M/year. Unnamed customers are labeled by what they rent, including $NVDA’s current B300 (Blackwell Ultra) and older H200 GPUs. $BRUN’s TCV is over $2.6B, so the other $989M+ is in deals without disclosed terms. The October 6 release calls the $2.6B “total revenue customers have committed to pay,” but the August 14 Q2 deck also counted renewal options and LOIs in TCV, and $BRUN hasn’t said whether the definition changed. So the conservative read counts only the seven contracts with disclosed terms, worth $1.61B, not the $2.6B headline.
@convequity 7 clicks

Convequity on X

Networking, not compute, now decides custom AI chip wins, and in our view that is Broadcom's ($AVGO) edge. As GPUs become more like specialized AI processors, matrix compute looks much the same everywhere. At 224G high-speed chip connections (SerDes), only Nvidia ($NVDA) and Broadcom can do it. If you want a custom chip, you go to Broadcom. That is why we avoid Marvell ($MRVL). A cheaper design partner is not necessarily cheaper overall: weaker connections and networking compound costs after deployment, resulting in higher total system costs. The bottleneck is moving from the switch to the optics. Lasers and optical signal-processing chips are short, and Broadcom can bundle priority optical supply with its custom chips, even for customers running Nvidia GPUs. Nvidia's lack of optical know-how is becoming its biggest disadvantage, and we expect Broadcom to take more share from it for years to come.
@convequity

Convequity on X

We see $NVDA seeking more power over its HBM suppliers by making Samsung, SK hynix and Micron ($MU) conform to the same base die design for NVHBM, its custom memory. Seoul Economic Daily reports Nvidia asked all three to make the NVHBM base die at TSMC ($TSM), citing consistent performance across suppliers and less validation work per supplier. Samsung reportedly agreed. The base die is the logic beneath the DRAM stack that connects it to the processor, and it is where each memory maker could differentiate with its own logic design. On NVHBM the base die follows Nvidia's standard and carries Nvidia's own memory controller. Our view: this makes the three suppliers' NVHBM stacks much closer to interchangeable. Nvidia can validate each supplier's memory more easily, shift orders between them and push harder on price. The trade-off is more reliance on TSMC. The memory makers are left competing on DRAM core dies, stacking, yield and capacity, which weakens their pricing power with Nvidia on NVHBM. What would change our view: SK hynix or Micron keeping their own logic in the NVHBM base die, or NVHBM staying a small share of Nvidia's memory purchases.
@gnoble79 24 clicks

George Noble on X

This is not what you want to see if you are a $CRWV bull. $CRWV remains one of our biggest shorts. We are short from $107. Only a takeunder from $NVDA will prevent bankruptcy. Subscribe to the Noble Update to receive our best ideas. 67% YTD performance based on a diversified book of longs AND shorts. We have unrivaled performance. We have decades of experience. We are battled tested and have been through numerous market cycles. We have a process. While others chase narratives and charts, we focus on valuation and fundamentals. KNOW WHAT YOU OWN. YOUR ATTENTION TO THIS MATTER IS GREATLY APPRECIATED.
@StableBread 11 clicks

Fajasy on X

I first wrote up an AI data center company that rents out $NVDA GPUs to AI startup companies on April 15, 2026, when it still traded as a SPAC at $10.70/share, just above its ~$10.60 trust value. The merger was a success, and the stock peaked at $39.20/share on June 1st. Now it trades below its SPAC price. That company is Boost Run $BRUN, which has fallen to $8.89/share, down 77.3% from its peak, even after it announced its largest disclosed contract on October 6 (worth $525.6M). The three larger listed GPU clouds have fallen from their 2026 highs as well, through the October 16 close: → CoreWeave $CRWV: Down 55.5% to $61.39/share. → $IREN: Down 60.2% to $26.99/share. → Nebius $NBIS: Down 42.3% to $165.47/share. So $BRUN fell with the group, but further than all three. I’d blame the drop from $39.20 on two causes that have nothing to do with $BRUN’s core operations: (1) higher rates and borrowing costs, and (2) $BRUN’s float growing ~3.1x, which was expected once its lockup ended and its warrants were called. Yet the business carries less risk than it did as a SPAC: → Contracts: $1.61B with disclosed terms and TCV over $2.6B, including $1.43B signed since my April write-up. → Customers: Thinking Machines Lab (the AI startup founded by former OpenAI CTO Mira Murati) and Cohere make up 61.9% of the $1.61B on three- to five-year terms, after three customers made up 75.9% of 2025 revenue, led by RunPod, a GPU cloud platform, at 44.5%. → Year-end target: If Thinking Machines’ GPUs are fully running by December 31 and the June order starts on November 30, the five disclosed contracts running by year-end cover $300.6M (75.1%) of the $400M ARR target. → Merger: Closed on May 8, 2026, with zero redemptions. → Balance sheet: The merger proceeds paid off its short-term bridge loans, the going-concern language is gone, and pro forma cash is $191.8M, up from the $108M I used in my June 3 update, right after the merger. But $BRUN still has to get the GPUs bought, financed, and running by December 31: → To reach the $1.0-1.2B of GPU spending management expects for 2026, $BRUN has to deploy $663-863M of GPU servers in H2 2026, 2.3-3.0x its cash and contracted prepayments. → CFO Erik Guckel said on the Q2 call that $BRUN is working on new ways to finance the build, but I found no new facility, lender, or term sheet in its filings or press releases from August 14 to October 16. → $BRUN’s named customers are private AI startups without investment-grade ratings, so its contracts are harder to borrow against. → If lessors funded 70% of the H2 2026 build instead of 79.5% (as in H1 2026), $BRUN would need $199-259M of its own cash for the build, more than its $191.8M of pro forma cash. → Both 10-Qs came late, after notices of late filing on May 26 and August 14, and $BRUN still reports three material weaknesses, the most serious kind of gap in a company’s financial reporting controls. → Founder and CEO Andrew Karos controls 87% of the vote. At $8.89/share, $BRUN’s EV is $989.7M. That’s 2.5x its $400M target for year-end ARR, down from 5.4x at the $22.93 price in my August 16 update. At 2.5x, $BRUN trades below all three peers, including $IREN at 3.2x. So did the thesis break, or just the stock? That’s what I answer in my new ~5k-word update on $BRUN, including my bear, base, and bull cases (Excel model included).
@convequity 7 clicks

Convequity on X

Nvidia ($NVDA) faces a major issue: most AI compute is moving to inference, where its CUDA moat protects less. CUDA's programming flexibility is essential for training, which is complex and changes rapidly. Inference runs a trained model with fixed weights, and we think inference chips are easier to build than training chips. The mix is shifting. Post-training RL, reasoning tokens and final replies are all forms of inference. On our view, pre-training could fall below 20% of compute in the future. In February, Cursor said the compute used to post-train its Composer 1.5 model surpassed the compute used to pre-train the base model. At Convequity we think CUDA's depth now protects a shrinking share of the workload, and we expect that to continue. Deeper dive available at Convequity.
@ACapitalLP 53 clicks

A Capital on X

𝙒𝙝𝙖𝙩 𝙝𝙖𝙥𝙥𝙚𝙣𝙚𝙙 𝙩𝙤 𝙩𝙝𝙚 𝙉𝙚𝙤𝙘𝙡𝙤𝙪𝙙 / 𝘼𝙄/𝙃𝙋𝘾 𝙩𝙧𝙖𝙙𝙚 𝙡𝙖𝙨𝙩 𝙬𝙚𝙚𝙠: I like distilling complex topics down to plain English. Here’s my understanding of what happened. Please correct where wrong… $CRWV and its peers rent $NVDA GPUs to AI labs and hyperscalers under multi-year take-or-pay contracts. The capex is so large relative to cash flow that the model only works if they can keep raising tens of billions a year in new financing to build what customers have already signed for. Much of that comes from secured loans backed by specific chips and contracts, while the rest comes from unsecured bonds, converts or stock. Importantly, not all secured debt is non-recourse, and many of $CRWV’s GPU loans are guaranteed by the parent. Still, the secured facilities held up considerably better than the unsecured bonds, which went from ~12% to ~15% yields, while the cost of insuring them via CDS roughly doubled… Nothing fundamental appears to have caused it. No major customer walked, GPU pricing held up, and the quarter looks fine. What happened was a cluster of headlines landing on a nervous market with the 10-year near 5.2% and hundreds of billions of AI debt issuance expected: Chamath calling it a private-credit bubble, an analyst saying the equity could go to zero, Gartner forecasting that over 80% of neoclouds stop operating independently by 2030, a few smaller peers failing to get IPOs and debt deals done. Underneath sat the bigger question of whether neoclouds are durable businesses or really just financing wrappers that hyperscalers eventually insource and cheap open-source models undercut. Once investors doubt the parent can fund growth, the stock falls, which makes the equity leg more dilutive, which worries bondholders more. Credit and equity fed on each other and the whole group fell 10-30% in sympathy. Was it justified? In three parts. (1) The $CRWV unsecured bond move was likely overdone: a company with ~$17bn of liquidity and a $129bn contracted backlog doesn’t suddenly become a distressed credit because of a handful of headlines. But backlog isn’t cash flow, and secured lenders have first claim on specific assets and cash flows, with many loans also carrying parent guarantees. The question for unsecured creditors is how much cash actually makes it back to the parent after servicing the secured debt. Still, a move to ~15% yields looks more like aggressive HF de-risking with headlines as the accelerant than a sudden collapse in the underlying business. (2) The $CRWV equity move was more understandable, just not for the headline reasons: this is a levered financing business whose equity returns depend on the cost of the marginal dollar of capital. When financing gets more expensive, the residual value on every future GW shrinks, and any gap that needs to be filled with equity becomes more dilutive. That said, unsecured bonds trading at 15% doesn’t mean every new GPU deployment costs 15% to finance. A 20% haircut for a stock that may not self-fund until 2030 could be seen as fair, but it ultimately depends on how much parent-level capital is actually required. (3) The group ($CORZ, $CIFR, $WULF) selloff was the least justified: $CRWV finances depreciating GPUs and bears the risk of obsolescence and replacement. The miners converting to HPC are increasingly financing long-lived power and data center infrastructure leased to hyperscalers under 15-20 year contracts, often with the customer owning the GPUs. Their risks are construction, power delivery and financing, not GPU replacement or compute pricing. Different assets, different contracts, different capital structures. Yet the market sold them as though they had the same problem, and that’s where the mispricing most likely sits…
@convequity 15 clicks

Convequity on X

We think the gross margin gap between Cerebras ($CBRS) and NVIDIA ($NVDA) says little about which company delivers compute more cheaply. Cerebras reported a 14% gross margin in Q2 2026, against roughly 75% for NVIDIA. That is not a clean comparison, because Cerebras's commercial agreements include customer warrants. A customer warrant gives a buyer of Cerebras compute the right to buy Cerebras shares at an agreed price, adding potential equity upside to the compute it purchases. Cerebras records the value of these incentives over time as a reduction in revenue. Its cost of delivering the compute stays the same, so the full amount comes out of gross profit and its reported gross margin falls too. Add the warrants and a few other items back and Cerebras's Q2 margin was 41%, so the reported gap tells us little about what each company spends to deliver equivalent compute. Cerebras could challenge NVIDIA with attractive economics on selected workloads without matching every capability of NVIDIA's platform. In our view NVIDIA's stronger answer is Feynman reaching customers on schedule with the 3D-stacked SRAM we expect. Stacking more fast memory close to the compute would reduce the need to fetch data from slower external memory, helping NVIDIA narrow the architectural advantage behind Cerebras's fast inference. NVIDIA's adoption of Groq technology could ease some of this pressure, but we believe Groq's current architecture remains weaker than Cerebras's for low-latency inference. The longer Feynman takes to arrive at scale, the more time Cerebras has to win customers.
@convequity 3 clicks

Convequity on X

The market still prices $NVDA's moat on one idea: the GPU is special because it can do anything. But Nvidia has spent every generation trading that flexibility for raw AI horsepower. Its chips now look less like general-purpose GPUs and more like the AI-specific chips Google and Huawei build. That's great for performance, but it chips away at the very thing that set Nvidia apart. The closer its GPU gets to a TPU, the easier it becomes to replace. Deeper dive available at Convequity.
@convequity 1 click

Convequity on X

3D chips could create more room for specialist EDA startups. We are not convinced that more difficult chip design automatically translates into a large revenue windfall for $SNPS, $CDNS and Siemens. Fine-pitch stacking changes decisions that used to be made within one die. Put a circuit upstairs and its wires may get shorter, but its heat affects the circuit below. Placement, signal timing and power delivery have to be optimized together across the stack. That creates a reason for $AVGO and $NVDA to build custom tools around their architectures. Broadcom already describes a proprietary design flow built on industry-standard tools. Nvidia publishes its own research on assigning connections across tiers in 3D chips. Our inference is that the chip designers will keep some of the most differentiating optimization work in-house. A startup that solves a costly gap in cross-die planning or optimization could also win business without replacing an entire EDA suite. The incumbents still supply implementation, verification and manufacturing sign-off tools, and they are developing 3D capabilities themselves. Internal tooling is evidence that customers need more than the standard flow; it does not establish that incumbent spending falls. We see a more interesting startup opportunity where a tool can shorten repeated design iterations or improve the resulting chip enough to earn a separate budget. That is the part of the 3D transition we would investigate before assuming the largest EDA vendors capture most of the upside.
@aleabitoreddit 3 clicks

Serenity on X

Glad we finally got a word too from $NVDA Jensen Huang today from an event at Stockholm... About $CRWV amid the broader Neocloud selloff. When asked about the markets reactions to Coreweave, Jensen responded: - "Coreweave's strength is improving every day, I'm optimistic" - "The cost of capital is actually decreasing" On whether Nvidia needed to provide a financial backstop: - Jensen said he did not see a need for Nvidia to backstop CoreWeave I'm personally not too positive about $CRWV relative to $NBIS and the others... but any copium helps from the man himself with market sentiment.
@convequity 1 click

Convequity on X

$NVDA's falling unit share in China's data-centre AI chip market overstates how much business Nvidia has lost to Huawei. Three things explain why. 1. IDC counts units, and did not say whether that count includes the Blackwell parts barred from China and reportedly smuggled in. Chinese demand also reaches Nvidia outside China. Xiaomi or ByteDance rents a Blackwell cluster in Malaysia from a local cloud provider. That is Nvidia revenue from a Chinese buyer that never appears as a Chinese unit. So Nvidia's share of what Chinese buyers spend on AI compute is higher than its share of counted units. 2. Huawei's volume growth is real, and memory explains how. Huawei cannot get enough stacked high-bandwidth memory, the HBM that top-end AI chips need. The 910C ran on an older grade, HBM2E, that Huawei had been able to mass procure. The new 950PR uses lower-bandwidth single stack memory instead of stacked HBM, so the 950PR does not depend on HBM supply. Each chip moves data more slowly, and Huawei can make many more of them. Also, we expect 950DT to ramp up quite quickly alongside 960DT coming earlier in 1Q27 instead of original 4Q27 timeline due to Huawei's own HBM variant that is able to ramp up production quicker. 3. So Nvidia's drop is not Nvidia failing to supply China. In our view, it has three causes: US export rules that bar its best parts, Beijing promoting domestic chips and requiring approval to import Nvidia's compliant ones, and a 950 series that beats the 910 series it replaces and is production-wise attractive for model serving. For Nvidia, where Chinese compute spend goes matters more than the China unit count, and much of it still goes to Nvidia hardware, in China or in Malaysia.

Investor Letters

Letters that discuss NVDA.

Every letter that discusses NVDA

Tickers on this page

Companies the stories above are also about, besides NVDA.