sf_lonely retweeted
Solid breakdown.
The real takeaway here is how quickly mechanical market dynamics can overwhelm fundamentals.
Between $ANT siphoning active liquidity away from the Mag7 and $AMD, the semi sell-off was almost an inevitable liquidity flush.
At the end of the day, fundamentals don’t matter in a degrossing.
🚨 Séance très violente sur Wall Street
Personne dans l'IA n'a envie de lire ce qui suit mais lisez-le quand même
Quelques explications :
1. Anthropic lève 81,3 Md$ et aspire la liquidité
IPO à 83 $, premier échange à 111,50 $, plus haut vers 115 $, vers 96 $ à 15h10. Plus de 2 000 Md$ de valorisation à l'ouverture, 80% des salariés deviennent millionnaires sur le papiers, dont 400 avec + de 100M.
La lecture qui domine est que pour acheter de l'ANT, les gérants ont vendu les Mag 7 et les semis. Pas d'entrée au Nasdaq-100 avant le 18 décembre, donc aucun achat indiciel pour compenser. $AMZN plus grosactionnaire, perd 3,9 %, Google pareil.
L'IPO se pase comme une IPO standard, open très haut, beaucoup de ventes de prise de profits, puis souvent une consolidation. Le consensus US était d'acheter la pré-ipo et la vendre directement, je ne suis donc absolument pas surpris.
2. Google accélère et des rumeurs folles d'AGI
Gemini 4 Neon est sorti hier : 61 sur l'index d'Artificial Analysis, 3e au classement, 3 $ le million de tokens. La rumeur parle d'un petit modèle entraîné avec de l'auto-amélioration récursive, et le DeepMind Institute publie un essai qui décrit l'AGI comme « à portée ». Un modèle aussi bon et moins cher pèse sur le prix des tokens, $GOOGL passe de +4,1 % à -2,8 % cependant
Google est connu pour faire du Benchmaxxing à l'inverse de Grok mais je ne vous cache pas que pour le coup là c'est extrêmement intéressant. Time to AGI : 0 ?
3. AMD a désormais son propre problème chinois
Hunterbrook se déclare short $AMD : IEIT, fabricant chinois de serveurs sur liste noire, aurait intégré au moins six versions récentes de son firmware.
Hunterbrook reconnaît n'établir aucune violation, AMD conteste. Le titre perd 9,4 %, le jour où la règle du Commerce sur Singapour (12,4 % du CA 2025) entre en vigueur. Ma lecture c'est que ça amplifie une news déjà bien salée, mais je ne pense pas qu'il y ait un problème sur AMD.
4. TSMC coupe une partie de la Chine, les semis rendent tout
TSMC et Samsung suspendent certaines livraisons à des clients chinois en attendant les licences du Commerce. Cela à surpris le $SOX de +4,7 % à -2,3 %, $NVDA de +4,75 % à -2,7 %, $MU de +7,8 % à -1,9 %. Nvidia file vers une sixième baisse d'affilée, $NBIS lâche 8,8 %, les neoclouds sont vraiment en enfer...
5. Le consensus shift doucement vers une capitulation
VIX vers 28, $QQQ sous son plus bas du 17 novembre. Les gens rangent les armes jusqu'à janvier, l'IPO d'Anthropic devait être un signal positif
Je ferait un post plus long pour expliquer ce que je retiens de mon voyage aux USA, après avoir parlé à des buyside & sell side. Meh.
6. Les rares îlots verts
Palo Alto +5,2 % après ses résultats et le rachat d'Applied Compute, $PDD +10,9 %. $BABA a touché +4,5 % avant de repasser dans le rouge. Le reste se fait juste enterrer.
7. Roaring Kitty est de retour
Il revient avec un meme Nintendo. Nintendo prend 15 % et $GME 7 % sur la news, hilarant mais je me tiendrai bien à l'écart
Demain je surveille $ANT face aux 83 $ de l'IPO, les suites sur $AMD et les licences pour TSMC, le KOSPI risque de déguster cette nuit
sf_lonely retweeted
BREAKING: 600 early Anthropic employees now worth over $100 million after the IPO
Over $1,000,000,000,000 wiped from stock market in past hour
sf_lonely retweeted
Replying to @DrAshokKMittal @NyaykiFitrat
Men are expected to be providers without ever running out of money, protectors without ever showing fear, and warriors without ever complaining. The moment they need support, they are left to fight alone.
Happy International Men’s Day!
sf_lonely retweeted
The biggest physical bottleneck in 800G & 1.6T AI optical networking right now is Indium Phosphide (#InP) laser capacity.
Globally, three Western giants lead the high-end InP race — and the shift to 6-inch wafers — each with a very different model:
1️⃣ $COHR (Coherent): First to ramp 6" InP globally. Massive merchant laser supplier + own transceivers. Direct beneficiary of laser scarcity pricing.
2️⃣ $LITE (Lumentum): Pure-play merchant leader in 200G/lane EML & high-power CW lasers (SiPh/CPO), aggressively expanding toward 6" capacity.
3️⃣ $NOK (Nokia): Operates an existing 3-inch InP fab in Sunnyvale alongside its new 6-inch InP fab in San Jose (Silicon Valley). While captive capacity shields Nokia's own systems from laser shortages, it also holds significant commercial upside to sell optical components & engines (like ICE-D) directly into AI data centers.
While $COHR and $LITE capture merchant scarcity margins, $NOK combines vertical supply security with merchant component upside.
#AI #OpticalNetworking #Photonics
sf_lonely retweeted
Replying to @Uppolice
@Uppolice
@varanasipolice
@AnilRajbharbjp
@DcpVns
@AcpSarnathVNS
Can you please help Senior citizens who are 90+ years old?
@jyotiTpandey05 @untoldStoryMens @kirankumargoli @Smash_Gynocracy @imra_mra @treeni @Sreekrish555 @SfLonely @KreatelyMedia @SKrishnaVijay @venkat_fin9
Even 93 yrs old Father
And 90 Yrs old mother is not safe......
Can a common citizen ask about it..what type of law, order and protection of Old age human being can be maintained. @adgzonevaranasi @deo_varanasi @varanasipolice this is ur constituency sir @narendramodi
sf_lonely retweeted
Replying to @LegalFightBack @Uppolice @varanasipolice @AnilRajbharbjp @DcpVns @AcpSarnathVNS @jyotiTpandey05 @untoldStoryMens @kirankumargoli @Smash_Gynocracy @imra_mra @treeni @Sreekrish555 @SfLonely @KreatelyMedia @SKrishnaVijay @venkat_fin9
Can you help 90+ years old elderly parents in distress? @DcpVns @deo_varanasi @varanasipolice @myogiadityanath @myogioffice
sf_lonely retweeted
Replying to @varanasipolice
@varanasipolice @DcpVns @AcpSarnathVNS @Uppolice Urgent police intervention requested. My wife has forcefully entered my residence in Sarnath, where my 93-year-old father and 90-year-old mother live peacefully.@realsiff @DeepikaBhardwaj
@SamSiff
@@deo_varanasi
@@khurpenchh
sf_lonely retweeted
I'm out of $AAOI. Honestly I had high hopes for $AAOI but it's got to a point where management can no longer be trusted. They have shown this through over-guiding, underdelivering, making comments on diversifying financing (and to their credit they secured a financing facility) only to drop another bombshell ATM on us shareholders.
But it's not just that they dropped another ATM. It's the fact that they chose to do it in this market where shares have fallen so damn low that they're going to have to issue way more shares than they did when they conducted the $600M ATM a couple months ago. It's the fact that that in this rugged, unfavorable tape, they have, without any reassurance to shareholders, rugged us with another ATM and we are well on our way to $40 in overnight markets.
I understand optics is a capex heavy business but then why did Thompson Lin say that spending for next year may ease then drop this news? Why did the CFO say they would rely on more debt backed financing then drop this news? It's this lack of delivery and communication that has led to widespread frustration not just for myself but in the wider $AAOI community. Honestly, business wise I believe the company is still set up to do alright next year if they hit estimates and growing pressure on Chinese optical component exports is still a real bullish tailwind for the company. I do think I probably sold a short term bottom here.
But you have to separate the business from the stock and concern yourself with how the market is shaping up. I don't think I can be a part of a company who seems to be the runt of the optics litter any longer.
Furthermore, I think we should all consider the opportunity cost of holding $AAOI when there are better positioned companies out there that are also on sale with more proven margins and revenue. On the other hand, Shareholders now not only have to bite the bullet on the ATM news but also have to go through the overhang that will be created by the dilution.
Pair that with the stock approaching lows with sentiment bottoming and you're stuck between a rock and a hard place. Even when we do bounce back up there are bagholders upon bagholders waiting to exit. And I am not certain carrying the beta of a growth story with execution risk in an unforgiving tape is wise at all when the next year is layered with uncertainty.
To me, cash right now is better than going down with a sinking ship imo or at least rotation into more quality names. I also have some concerns about the way the market is heading right now and how valuation compressions are happening across the entire AI sector. $AAOI is not the same stock it was in June nor is it the same type of market. Capital preservation is paramount now and therefore I will be exiting $AAOI.
sf_lonely retweeted
$NBIS
Water Tower Research after earnings:
2027 revenue: $11.7B → $12.45B
2027 capex: $26.0B → $42.5B
2028 revenue: $22.6B (new)
2028 EBITDA: $11.6B (new)
The note is titled "Cloudy" ARR. It still models revenue up ~7x from 2026 to 2028 (my math).
sf_lonely retweeted
🚨$NVDA cuts neocloud support plans down $1B q/q and adds tighter language around conditions; removes revenue share language from last quarter; adds new language that certain commitments "could be reduced or eliminated" if contingencies are not satisfactorily resolved. Cuts previously disclosed 2028 capex $1B to zero (sounds like a project pushout or cancellation).
Now says commitments take effect "only if the neoclouds secure financing to purchase the infrastructure products and the related compute capacity is deployed, made operational, and accepted by us."
Previous language: "We continue to make commitments....Under these agreements we will earn revenue on upfront sale of our infra and if certain criteria are met, we will participate in revenue share."
Re: secondary equity investments, future commitments fall from $25B to $20B but 2028 plans rise from $3B --> $11B. Unclear if this is pushouts from 2027 investments not made (such as Figure AI).
Bottom line: Total commitments grew 41% QoQ, driven almost entirely by a surge in 2028 supply/capacity commitments and long-dated data center leases, but the new contingency clause means a large portion is conditional on neoclouds securing external financing.
Also 10-Q to arrive "in coming weeks" seems like a delay.
sf_lonely retweeted
David Tepper’s 13F
I would not have guessed the Appaloosa’s portfolio is all-in on the AI trade.
One might be forgiven for confusing it with Leopold’s portfolio.
There is no China. No energy.
He rotated out of AMD into Nvidia.
The analysts there are AI pilled.
Tepper’s 13F is always one I look forward too.
These are interesting ideas… it’s just surprising to see Tepper whole hogging AI.
BTW, the Lumida Invest app imputed 13F returns so you can estimate how the funds performed and get their factor exposures.
Overall, brutal quarter for hedge funds.
$SPCX $MU $ORCL $NVDA $AMAT
sf_lonely retweeted
$AAOI management lost all credibility by launching another $550M ATM.
Especially after it comes contrary to earlier commentary about other financing methods + share price caution.
I still find AOI operationally compelling around their CPO program + optical transceiver ramp.
Especially as it comes to the new China optical transceiver restrictions and incoming $5B-7B LTAs.
But I misjudged management and suffered heavy losses after cost averaging up to $100+ from the initial ~$25.
Definitely did not foresee them tapping into abusive financing methods again after the sector drop.
I’ll still track their progress, but I don’t see how it’s possible to support a company that serially dilutes with ATMs.
sf_lonely retweeted
atlaspeakresearch.com/earnin…
$NVDA KEY READ-THROUGHS FROM NVIDIA’S Q3 FISCAL 2027 EARNINGS CALL
NVIDIA’s results strengthen the outlook for AI infrastructure spending, but the more actionable conclusion is that economic benefits are becoming increasingly uneven across the supply chain. The strongest positive signals favor scarce memory, qualified substrates, advanced packaging, optical components and already-operational computing capacity. The clearest negatives affect thin-margin hardware buyers, infrastructure projects awaiting power or commissioning, and Chinese frontier-model developers facing tighter access restrictions. The increase in fiscal 2028 revenue-growth guidance from 70% to 75% supports continued investment, while NVIDIA’s rising commitments and customer-support arrangements show that financing and deployment readiness increasingly determine who captures that growth. Magnitude assessments below distinguish business-level exposure from consolidated earnings sensitivity; they do not imply specific share-price returns. Country labels identify issuers’ home markets.
MEMORY: SUPPLIERS ARE CAPTURING ECONOMIC VALUE THAT NVIDIA IS TEMPORARILY ABSORBING (READ-THROUGH 1)
Affected companies: Micron Technology (MU: US) and SK hynix (000660: South Korea). Direction: Positive. Magnitude: High for memory earnings visibility and pricing durability, without sufficient disclosure to allocate NVIDIA’s commitments between suppliers.
The strongest evidence is not simply NVIDIA’s revenue growth. Its CFO commentary states that supply commitments increased from $279 billion to $390 billion, “primarily related to the multi-year procurement of memory supply.” Scheduled fiscal 2028 supply commitments increased from $87 billion to $203 billion. Meanwhile, NVIDIA reaffirmed a gross-margin decline to 71%-72% in Q4 and explained that it was honoring existing customer purchase orders before revised pricing takes effect. This combination indicates that a meaningful portion of the AI system’s economic value is accruing upstream to scarce components.
The transmission mechanism is stronger contracted demand, better production visibility, favorable product mix and pricing support for qualified memory suppliers. Micron’s disclosed Rubin-related products include HBM4 and SOCAMM2 memory for Vera platforms, extending its exposure beyond GPU-attached HBM. SK hynix’s October 28 results independently corroborate the contractual shift: agreements cover more than 50% of its 2027-2028 production capacity, with pricing mechanisms that reflect market conditions within negotiated ranges.
The near-term catalyst is increased confidence in the duration of memory pricing and contracted demand, rather than an assumption that all $111 billion of incremental commitments becomes immediate supplier revenue. The longer-duration benefit is improved utilization and investment visibility into 2027-2028. The principal limitation is that long-term agreements do not eliminate pricing cyclicality, qualification risk or eventual capacity oversupply. Nevertheless, this is a stronger signal for memory profitability than a generic increase in AI capital expenditure.
SUBSTRATES: CUSTOMER-SPECIFIC CAPACITY AGREEMENTS FAVOR IDENTIFIED SUPPLIERS, NOT THE ENTIRE GROUP EQUALLY (READ-THROUGH 2)
Affected companies: Kinsus Interconnect Technology (3189: Taiwan) and Samsung Electro-Mechanics (009150: South Korea). Direction: Positive. Magnitude: Potentially high for Kinsus’s NVIDIA-related substrate business; moderate at Samsung Electro-Mechanics’s consolidated level.
NVIDIA explicitly announced new substrate agreements with these companies. The accompanying November 17 supplier release states that Kinsus will expand advanced ABF substrate capacity dedicated to NVIDIA to more than 3.5x current levels. Production is expected to begin in the second half of calendar 2027, with the additional capacity targeted for Q3 2028. Samsung Electro-Mechanics amended its existing agreement to improve production capabilities at existing factories.
These are identifiable commercial relationships rather than thematic exposure to semiconductor demand. Larger and more complex computing packages require qualified substrates with demanding routing, dimensional-stability and reliability characteristics. Dedicated capacity arrangements improve demand visibility and support capital investment in capabilities that cannot be substituted immediately.
The distinction between existing-factory improvements and greenfield expansion matters. Samsung Electro-Mechanics potentially has a shorter operational path to incremental output, while Kinsus has a larger disclosed dedicated-capacity expansion but a later realization schedule. The 3.5x figure applies to NVIDIA-dedicated capacity, not Kinsus’s total revenue or earnings.
The immediate trading catalyst is stronger confidence in customer allocation and the durability of advanced-substrate demand. The fundamental catalysts are equipment installation, customer qualification and utilization through 2027-2028. Initial depreciation, yield learning and ramp costs can precede meaningful earnings contributions. Treating the announcement as an immediate 3.5x earnings uplift would materially overstate its significance.
FOUNDRY AND PACKAGING: NVIDIA IS FINANCING BOTTLENECK RELIEF, NOT DEMONSTRATING FOUNDRY DISINTERMEDIATION (READ-THROUGH 3)
Affected companies: Amkor Technology (AMKR: US), ASE Technology Holding (3711: Taiwan), through wholly owned subsidiary Siliconware Precision Industries, and Taiwan Semiconductor Manufacturing Company (2330: Taiwan). Direction: Positive. Magnitude: High strategic significance for Amkor and SPIL’s advanced-packaging businesses; positive but less directly quantifiable at the consolidated TSMC and ASE levels.
The call named SPIL and Amkor among NVIDIA’s expanding packaging partners. NVIDIA’s accompanying release specifies nearly $3 billion of previously announced Amkor agreements supporting Arizona and South Korean capacity, including prepayments. The new SPIL agreement includes multibillion-dollar funding and targets additional contracted advanced-packaging capacity by mid-2029.
The transmission mechanism combines future service revenue with reduced external funding requirements. Customer prepayments can help suppliers build capacity before production revenue arrives, improving project financing and visibility. They are not immediate packaging revenue, and the nearly $3 billion Amkor figure includes agreements announced in July and October rather than an entirely new award on this call.
The less obvious conclusion is that broader outsourced packaging participation does not establish a loss of TSMC’s advanced-node wafer business. NVIDIA’s November 17 release explicitly identifies TSMC’s 5nm, 3nm and CoWoS technologies in Vera Rubin and Vera CPU shipments. Additional assembly, interposer and testing capacity can relieve constraints around a growing platform without displacing the underlying foundry relationship.
Near-term catalysts are funding, construction and qualification milestones. The largest disclosed SPIL packaging contribution is a 2028-2029 development rather than a near-term quarterly earnings event. Geographic diversification into the US and South Korea also has strategic value, although higher costs and ramp inefficiencies can delay profit conversion.
OPTICAL COMPONENTS: CO-PACKAGED OPTICS SHIFTS VALUE TOWARD LASERS AND INTEGRATION RATHER THAN ELIMINATING OPTICS DEMAND (READ-THROUGH 4)
Affected companies: Lumentum Holdings (LITE: US) and Coherent (COHR: US). Direction: Positive. Magnitude: High strategic relevance to their data-center optics franchises; near-term earnings magnitude depends on product mix and shipment timing.
NVIDIA reported networking revenue exceeding $21 billion, co-packaged-optics scale-out products in volume production and a production deployment connecting more than 10,000 GB300 GPUs. Management cited 5x better power efficiency, 15x higher reliability and 10% lower latency versus pluggable optics in the referenced configuration.
The relevant supplier relationships are established. NVIDIA’s March announcements with Lumentum and Coherent included multibillion-dollar purchase commitments, capacity-access rights and $2 billion investments in each company. The agreements are nonexclusive. The current call therefore provides evidence of commercialization against an existing investment framework, rather than announcing those commitments again.
The mechanism is greater demand for advanced laser components, optical networking products and package integration as bandwidth moves closer to switching silicon. Integrating optical functions does not remove the need to generate, transmit and receive optical signals. Consequently, a blanket negative view on optical suppliers because of CPO adoption would overlook the retained and potentially expanding upstream component opportunity.
The near-term catalyst is evidence that CPO has progressed into production deployment. The longer-duration catalyst is broader adoption across 2027-2028 infrastructure generations. The call does not establish supplier allocations within the Lambda deployment or justify applying NVIDIA’s networking growth directly to either supplier’s revenue. It also does not support a blanket short of companies with pluggable-optics exposure, since product portfolios and retained content differ.
CPU ARCHITECTURE: AGENTIC COMPUTING EXPANDS ARM’S OPPORTUNITY BEYOND GPU HOSTING (READ-THROUGH 5)
Affected company: Arm Holdings (ARM: UK). Direction: Positive for the architecture and royalty opportunity. Magnitude: High strategic importance, but unquantified and likely more gradual at the consolidated earnings level.
NVIDIA stated that agentic AI is driving CPU demand “far outstripping supply” and expects total CPU revenue to more than double in fiscal 2028. Newly named Vera adopters included CoreWeave, Meta, Nscale and OpenAI, joining AWS, Oracle Cloud Infrastructure and other previously announced customers.
Arm’s own technical commentary confirms that Vera uses NVIDIA-designed Olympus cores built on the Arm architecture. The mechanism is broader deployment of Arm-based processors for agent execution, code interpretation, retrieval, orchestration and scientific workloads. These functions generate CPU demand alongside accelerators rather than simply replacing CPU work with GPUs.
The important distinction is between NVIDIA’s CPU revenue and Arm’s monetization. Arm does not receive the full selling price of a Vera processor, and custom Olympus cores should not be confused with NVIDIA purchasing a complete standard Arm CPU design. The opportunity is architecture licensing and royalty-bearing shipments under the applicable agreements, whose economics were not disclosed on the call.
The near-term catalyst is adoption and shipment evidence supporting a larger data-center royalty base. The longer-duration shift is an expanding software ecosystem around Arm-based agent infrastructure. The evidence supports this architecture-level positive more strongly than it supports a quantified earnings short in Intel or AMD: additional CPU demand can expand the market even as suppliers compete for workload allocation.
CONSUMER HARDWARE: THE SAME MEMORY CYCLE THAT BENEFITS SUPPLIERS CREATES A PRICING-AND-VOLUME TRAP FOR PC OEMS (READ-THROUGH 6)
Affected company: HP Inc. (HPQ: US). Direction: Negative for underlying PC unit demand and margin flexibility. Magnitude: Material because Personal Systems is a large, thin-margin business, although the call does not establish an incremental earnings miss.
NVIDIA’s CFO commentary states that higher memory and system prices continue to affect consumer PC demand. This is particularly informative because NVIDIA simultaneously reported strong workstation and physical-AI demand. Enterprise infrastructure strength should therefore not be interpreted as evidence of a uniform hardware recovery.
HP’s latest reported quarter provides direct corroboration. Personal Systems revenue increased 18%, but units declined 16%, while operating margin was 4.6%. The divergence indicates substantial price-and-mix effects rather than healthy unit expansion across the portfolio. HP also reported improved memory supply and raised guidance, so NVIDIA’s commentary should not be treated as proof that HP’s procurement mitigation has failed.
The transmission mechanism is unfavorable regardless of the immediate response: absorbing higher memory costs compresses margins, while passing them through can pressure replacement demand and purchasing configurations. At a 4.6% segment operating margin, an unrecovered 100-basis-point cost increase would mechanically reduce segment operating profit by approximately 22%, assuming unchanged revenue and other costs. This is a sensitivity, not a forecast.
The near-term catalyst is further evidence on component pass-through, units and premium mix. The longer-duration risk is that AI infrastructure procurement sustains memory tightness beyond a normal consumer replacement cycle. The negative read-through is to underlying economics, not an automatic reduction to guidance that already incorporates significant cost pressure.
NEOCLOUDS: HIGH RENTAL PRICES IMPROVE THE ECONOMIC LIFE OF EXISTING GPU FLEETS (READ-THROUGH 7)
Affected company: CoreWeave (CRWV: US). Direction: Positive for installed-fleet revenue durability and asset recoverability. Magnitude: High operating relevance, with a substantially more qualified implication for equity cash flows.
NVIDIA reported that certain Hopper contracts had renewed at up to 2x their original prices, B300 rental rates had risen more than 60% since introduction, and its compute remained fully utilized across the hyperscalers and neoclouds it serves. These observations challenge the assumption that a new architecture necessarily makes older GPU fleets economically obsolete immediately.
CoreWeave’s November 10 presentation reported more than 1.7 GW of active power and confirmed that Rubin systems were already running production workloads. Operational capacity is the critical exposure: existing sites can monetize scarce compute without waiting for the full construction cycle required by new developments.
The transmission mechanism is stronger renewal economics, sustained utilization and potentially greater residual value after initial contracts expire. These factors can improve lifetime asset returns and lender recovery assumptions. However, the quoted rental increases apply to particular contracts and platforms, not CoreWeave’s entire fleet. Fixed-price customer contracts reprice only when their terms permit.
The near-term catalyst is evidence of renewals and utilization holding up despite Rubin’s ramp. The longer-duration benefit is reduced economic-obsolescence risk. The qualification is substantial: CoreWeave reported a 62% adjusted EBITDA margin but only a 10% adjusted operating margin and an adjusted net loss in Q3. Strong equipment-level economics do not remove depreciation, interest expense or future financing needs. NVIDIA’s support arrangements should not be interpreted as an unconditional guarantee of CoreWeave’s obligations.
NEOCLOUDS: ADDITIONAL GPU SUPPLY DOES NOT SOLVE POWER AND COMMISSIONING DELAYS (READ-THROUGH 8)
Affected company: Nebius Group (NBIS: Netherlands). Direction: Negative for assumptions of immediate revenue acceleration; not evidence of weakening underlying demand. Magnitude: High for near-term revenue timing and capital efficiency.
NVIDIA’s allocation criteria explicitly require completed land, power and shell, available capital and credible offtake. Kress emphasized that customers must be ready to use shipped infrastructure efficiently. The CFO commentary separately states that relevant neocloud commitments activate only after financing is secured and capacity is deployed, operational and accepted.
Nebius’s November 12 shareholder letter demonstrates why that distinction matters. The company moved expected 2026 revenue to the low end of its $3.0 billion-$3.4 billion range, increased capital-expenditure guidance to approximately $27 billion and shifted its expected year-end revenue run-rate milestone into February. It identified delays at third-party sites, including power-related work in New Jersey, while stating that the customer contracts remained in place.
The mechanism is a mismatch between capital deployment and revenue activation. More readily available Rubin systems can increase hardware readiness without resolving electrical infrastructure, commissioning or acceptance dependencies. NVIDIA can therefore raise its supply-supported outlook while an individual customer’s revenue remains delayed.
The near-term catalyst is site energization and customer acceptance, not another increase in industry GPU availability. The longer-duration benefit depends on converting owned-site construction into operational capacity. Importantly, Nebius also disclosed that prepayments on qualifying Q3 agreements exceeded associated capital expenditures. That is counterevidence to a blanket funding-crisis thesis. The high-conviction negative is timing and utilization of invested capital, not customer cancellation or an inference that all future spending is unfunded.
HYPERSCALE CLOUD: RUBIN DEPLOYMENT SUPPORTS REVENUE CONVERSION, BUT DOES NOT GUARANTEE BETTER FREE CASH FLOW (READ-THROUGH 9)
Affected companies: Amazon.com (AMZN: US) and Microsoft (MSFT: US). Direction: Positive for cloud revenue capacity; negative for near-term capital intensity. Magnitude: High within AWS and Azure, with differing consolidated cash-flow consequences.
NVIDIA reported 23% sequential hyperscale revenue growth, approximately 1-month Rubin production deployment at hyperscale customers and cloud-industry backlog of approximately $2.5 trillion. Management also described customer cloud revenue growth accelerating for the 6th consecutive quarter.
The mechanism is faster conversion of contracted demand into usable cloud capacity. Shorter deployment cycles reduce the period during which purchased hardware generates no service revenue. For customers constrained by available infrastructure, additional productive capacity can support growth without requiring a comparable increase in customer acquisition spending.
The peer evidence supports the revenue conclusion. Amazon reported 39.3% AWS growth in its October 29 release. Microsoft’s October 28 call reported 48% Azure growth, with GPU efficiency gains helping it allocate more capacity to customers despite third-party delivery delays. NVIDIA’s call corroborates those already reported results and provides a forward deployment signal; it is not a prediction of those completed quarters.
The cash-flow conclusion is less uniformly positive. Amazon reported a trailing-12-month free-cash-flow outflow of $42.4 billion, whereas Microsoft generated $18.6 billion of quarterly free cash flow. These periods are different, but the disclosures illustrate materially different funding outcomes. The near-term catalyst is capacity becoming billable. The longer-duration test is whether utilization, pricing and asset productivity outrun depreciation, power costs and continuing capital expenditure. NVIDIA’s sales growth alone cannot answer that question.
CHINESE INTERNET AND CLOUD: REMOTE-ACCESS CONTROLS ARE MORE DIRECTLY NEGATIVE FOR FRONTIER MODELS THAN FOR ALL EXISTING INTERNET EARNINGS (READ-THROUGH 10)
Affected companies: Alibaba Group Holding (9988: China) and Tencent Holdings (0700: China). Direction: Negative for access to covered advanced computing and frontier-model development. Magnitude: Potentially high strategically, but the immediate consolidated earnings effect is narrower and unquantified.
The call disclosed $2.4 billion of Q3 licensed Hopper shipments to China, no further such shipments and no China Data Center compute revenue in guidance. More importantly, management acknowledged that it was still examining the newly announced export controls. A lack of direct NVIDIA shipments to China does not capture the full exposure when Chinese companies obtain computing services through overseas data centers.
The public-inspection text of the new rule extends controls to covered remote computing supplied to restricted-country companies and adds destination-based licensing requirements for Indonesia, Malaysia, Singapore, Thailand and Brazil, subject to specified exceptions. Alibaba and Tencent have been reported to use advanced US computing through overseas facilities. The mechanism is greater licensing friction, potential capacity disruption, additional compliance requirements and pressure to redesign workloads around available domestic alternatives.
The important limitation is the rule’s conditional treatment of exclusively commercial recommender systems. Its definition excludes systems generating novel text, images, video, software or other content. Consequently, a frontier generative-model program and a narrowly defined existing recommendation workload should not receive identical earnings assumptions. Treating the rule as an immediate shutdown of all advertising or e-commerce computing would overstate the evidence.
The near-term catalysts are the scheduled November 19 publication, licensing interpretation and service-provider responses, including transition provisions. The longer-duration effect is a wider constraint on frontier-model development and a stronger incentive to build domestic computing ecosystems. US-parent exceptions for certain destination controls do not exempt US cloud operators from restricted-end-user screening.
POWER AND COOLING: FACTORY-BUILT AI INFRASTRUCTURE INCREASES THE VALUE OF QUALIFIED ELECTRICAL AND THERMAL INTEGRATION (READ-THROUGH 11)
Affected companies: Vertiv Holdings (VRT: US) and Schneider Electric (SU: France). Direction: Positive. Magnitude: High strategic relevance to their data-center businesses; moderate to high potential revenue significance, without disclosed supplier allocations.
NVIDIA introduced DSX POD as a modular AI-factory architecture, with initial shipments planned for early 2027. The call described moving cluster integration into the supply chain and reducing deployment from months to weeks. NVIDIA’s detailed announcement identifies Vertiv and Schneider Electric as partners and specifies that rack-level power, liquid cooling, piping, wiring and validation are integrated before delivery.
The mechanism is broader factory-integrated equipment content, repeatable designs and earlier procurement of qualified electrical and thermal systems. These suppliers address a bottleneck that determines when expensive compute begins generating revenue. Standardization can expand project throughput and improve repeatability, although competitive tenders and execution still determine supplier margins.
The critical distinction is that DSX POD reaches a prepared site. It does not eliminate grid interconnection, generation availability, permitting or site-level infrastructure. Similarly, Rubin’s claimed 10x improvement in tokens per watt does not imply a 90% decline in aggregate power-equipment demand. Under a constrained power envelope, customers can use efficiency gains to produce more output while continuing to deploy the available electrical capacity.
The near-term catalyst is conversion of the reference architecture into actual partner orders and early-2027 deliveries. The longer-duration shift is toward standardized, factory-tested infrastructure rather than bespoke integration at each site. This supports identified equipment suppliers more directly than it supports a blanket positive view on utilities or a blanket negative view on field contractors, many of which retain installation and commissioning work.
AI SYSTEMS AND CONTRACT MANUFACTURING: FASTER FACTORY INTEGRATION SUPPORTS VOLUMES, BUT THE GPU INVOICE IS NOT THE SUPPLIER’S ECONOMIC VALUE ADDED (READ-THROUGH 12)
Affected companies: Dell Technologies (DELL: US) and Hon Hai Precision Industry (2317: Taiwan). Direction: Positive for AI-system demand and manufacturing utilization. Magnitude: High within the relevant AI infrastructure businesses, but more moderate and execution-dependent for consolidated earnings.
The call identified Dell and Foxconn among DSX POD partners, reported nearly $20 billion of initial-quarter Rubin revenue and maintained annual architecture transitions. NVIDIA’s accompanying supplier release also described Foxconn’s expanded role across next-generation GPU and networking platforms.
The mechanism is greater demand for rack assembly, cluster integration, cabling, testing, thermal integration and delivery logistics. Factory validation can reduce customer deployment friction and make manufacturing reliability a more important competitive differentiator. Dell’s enterprise
sf_lonely retweeted
Muddy Waters is short $NVDA
A woman identified as Cenly Chen, referred to in some photograph tags as Cenly Puri, has been photographed in public between 2024 to 2026 with Ajay K. Puri. Until this June, Mr. Puri was likely the third most senior executive at NVDA, behind only Jensen Huang and Colette Kress. His title was EVP Worldwide Field Operations and his total compensation was slightly above Ms. Kress’ in fiscal years 2024 to 2026.4 Mr. Puri was one of only five NVDA executives named in the 2026 Proxy.
Separately, consolidated class action complaint filed against Super Micro Computer Inc. in October 2026 alleges that former SMCI Chief Growth Officer, a woman named Cenly Chen, “managed [SMCI’s] acquisition of GPU chips from NVIDIA” and “decided how to distribute [NVDA GPUs] among Super Micro’s customers.” In March this year, the U.S. indicted three individuals associated with SMCI, including a co-founder, for illegally selling at least approximately $2.5 billion of SMCI servers with NVDA GPUs to end customers in China through a Southeast Asian front company. Referring to the ex-SMCI executive, the class action complaint alleges that there was “no way Cenly Chen lacked knowledge of [SMCI’s] illicit sales.”
Mr. Puri left NVDA in late June 2026, and the former SMCI executive named Cenly Chen left SMCI in late April or early May 2026. Mr. Puri continues to serve in a senior advisory role to NVDA.
Nobody by the name of Cenly Chen has been indicted or otherwise accused of criminal conduct by any government. The allegations in the class action are unproven.
The question is whether the woman identified as Cenly Chen or Cenly Puri in photographs with Mr. Puri is the same Cenly Chen formerly employed by SMCI.
sf_lonely retweeted
🤝 #NVIDIA CEO Jensen Huang reportedly reaffirmed #SKhynix as its largest memory supplier, expected to remain so going forward. The partnership now extends beyond #HBM4 in Rubin to Vera CPUs, which feature LPDDR5X memory.💡More: pse.is/96hhbg 🔗
sf_lonely retweeted
A pivotal week looms for #NVIDIA—and its chip suppliers. With Jensen Huang out and the CFO taking the helm for now, will cost-cutting take center stage? Even Trump has floated Tim Cook as a successor. For #Samsung and #SKhynix, the bigger question is what comes next for HBM orders and pricing.💡More: pse.is/9v6h87 🔗
sf_lonely retweeted
13F REPORTS
Atreides Management (Gavin Baker) Q3 2026
Reported 13F exposure rose 25.5% to $18.0B from $14.3B, driven mostly by options.
- Option underlying value: $7.3B from $3.2B (+127.6%)
- Common stocks: $10.7B from $11.2B (-3.6%)
NEW STOCKS
- Figma $FIG: $269.2M
- On Holding $ONON: $192.2M
- Carvana $CVNA: $155.4M
- Microsoft $MSFT: $141.5M
- Autodesk $ADSK: $46.1M
- TTM Technologies $TTMI: $26.5M
- Hut 8 $HUT: $23.3M
BIGGEST ADDS
- Nebius $NBIS: +1,660%
- Meta $META: +116.6%
- Akamai $AKAM: +109.6%
- Astera Labs $ALAB: +102.1%
- Ciena $CIEN: +50.9%
- Nvidia $NVDA: +18.2%
- Cerebras $CBRS: +12.1%
- Samsara $IOT: +6.6%
BIGGEST CUTS
- Tesla $TSLA: -61.5%
- Micron $MU: -57.3%
- Amazon $AMZN: -41.6%
- Nokia $NOK: -41.4%
- Palo Alto Networks $PANW: -34.0%
- Credo $CRDO: -25.0%
- Unity $U: -24.5%
- Affirm $AFRM: -23.9%
- GitLab $GTLB: -16.2%
- CoreWeave $CRWV: -4.6%
- Coherent $COHR: -2.4%
EXITS (Q2 values)
- Alphabet $GOOGL: $174.5M
- Amphenol $APH: $112.9M
- Semtech $SMTC: $89.2M
- Cipher Digital $CIFR: $87.9M
- Dick's Sporting Goods $DKS: $85.1M
- ACV Auctions $ACVA: $71.2M
- Wix $WIX: $66.2M
- Mastercard $MA: $48.1M
- ServiceTitan $TTAN: $38.3M
- Ferguson $FERG: $35.1M
- Natera $NTRA: $28.3M
- Warby Parker $WRBY: $21.2M
OPTIONS (underlying value)
- New: semiconductor ETF $SMH puts ($1.2B), Alphabet calls ($688.2M), Unity calls ($196.5M)
- Bigger: Nasdaq 100 ETF $QQQ puts (+59.4% to $3.8B), Meta calls (+58.3% to $1.4B)
- Closed: Cerebras calls
BIGGEST HOLDINGS
- SpaceX $SPCX: $4.1B
- QQQ puts: $3.8B
- Meta calls: $1.4B
- SMH puts: $1.2B
- Astera Labs: $875.4M
sf_lonely retweeted
GAVIN BAKER BOUGHT FIGMA AND DOUBLED META
Gavin Baker's Atreides opened a $269M stake in Figma $FIG in Q3, more than doubled Meta and Astera Labs, and sold out of Alphabet and Amphenol, while cutting Micron by 57%. SpaceX is still its largest holding at 23% of an $18.0B U.S.-listed portfolio.
Biggest share increases:
Nebius +1,663%
Meta +117%
Akamai +110%
Astera Labs +102%
Ciena +51%
Biggest share reductions:
Micron -57%
Amazon -42%
Palo Alto Networks -34%
Credo -25%
Unity -25%
GitLab -16%
New positions:
Figma $269M
On Holding $192M
Carvana $155M
Microsoft $142M
Autodesk $46M
TTM Technologies $26M
Hut 8 $23M
Completely exited:
Alphabet
Amphenol
Semtech
Cipher Digital
Dick's
ACV Auctions
Interestingly, Atreides raised its Nebius stake 17-fold while adding new positions in Microsoft, On Holding and Carvana.
Holdings as of Sept. 30, 2026.













































