Straggler Liu | AI & Semis

SpaceX's AI Expansion: The Bottleneck Has Moved Beyond GPUs SpaceX's latest AI infrastructure plans reveal a shift in the economics of scaling compute. Musk expects 2.5–3 GW of operational capacity by year-end 2026 and targets roughly 13 GW by the end of 2027. But the more revealing number is 20 GW of planned power and cooling infrastructure. The constraint is no longer just acquiring GPUs. It's getting entire systems powered, cooled and earning revenue. Three developments matter. First, financing is becoming a competitive advantage. SpaceX's announced $40B Apollo-backed Vera Rubin leasing arrangement could accelerate deployment without requiring equivalent upfront cash expenditure. But leasing changes the timing of payments, not the underlying economics of hardware depreciation. Second, payback needs a denominator. Management claims sub-one-year payback on certain AI infrastructure investments, potentially falling to six months. That is extraordinary if repeatable. But does it include generation, cooling, buildings, financing and replacement costs? Accelerator-level payback and full-system returns are very different measures. Third, the bottleneck doesn't automatically determine who captures the profits. NVIDIA has clear direct exposure through planned GB300 and Vera Rubin deployments. Vertiv and Bloom Energy are credible infrastructure beneficiaries, but neither has a confirmed SpaceX supply relationship in the evidence reviewed. Their investment cases depend on pricing power and realized margins, not simply rising electricity demand. The 2027 capacity target remains a management ambition, not secured operational capacity. Orbital compute is a longer-term option, not a near-term solution to terrestrial power constraints. My investment thesis: AI infrastructure winners will be determined less by how much compute they finance than by how quickly they turn it into profitable, operational capacity. The numbers to watch: commissioned GW, utilization, full-system cash payback and financing costs. A financed GPU is not yet a productive GPU. $SPCX

Semiconductors Look Cheap on 2028 Earnings. But Which Earnings Will Survive? Morgan Stanley reportedly favors AI compute and memory after the selloff, with Nvidia and Broadcom among its preferred names. The fundamentals are substantial. Nvidia reported 75% gross margins last quarter. Broadcom generated $13.7B in quarterly free cash flow. Micron's FY2026 GAAP gross margin reached 80.7%. But that creates a valuation problem. A low multiple on exceptionally high earnings isn't necessarily cheap. Memory is particularly interesting. Sandisk says new customer agreements will cover roughly two-thirds of its FY2028 bit volume, with minimum financial guarantees and structured pricing. That could make NAND earnings more durable than in previous cycles. The question is how much protection those contracts provide when supply catches up. For compute, the equivalent test is whether margins and cash generation hold as AI infrastructure financing becomes more selective. My hypothesis: the next semiconductor winners won't simply be those with the lowest 2028 P/E. They'll be the companies whose earnings survive the assumptions embedded in that P/E. $NVDA $AVGO

More DRAM Wafers in 2027 Doesn't Necessarily Mean More Commodity DRAM Supply. Samsung, SK hynix and Micron are expanding capacity. But fab timelines and wafer-start estimates tell only part of the story. Micron has disclosed a roughly 3:1 HBM-to-DDR5 capacity trade ratio, with that ratio increasing in newer HBM generations. That means HBM growth can absorb a disproportionate share of new DRAM capacity. The counterweight is process migration: more usable bits per wafer could offset some of that pressure. My hypothesis is that investors focusing on rising wafer starts may overestimate conventional DRAM supply growth. The numbers to watch are HBM capacity allocation, effective bit output per wafer and actual commodity DRAM bit shipments. More wafers can coexist with tighter conventional DRAM supply. The mix matters more than the headline capacity.

AI Chipmakers' CDS Spreads Have Overtaken Hyperscalers'. But Who Actually Owns the Risk? Goldman Sachs trading-desk data reportedly show a basket of Nvidia, Broadcom, AMD and Intel CDS spreads rising to roughly 160bp, above a basket of major AI infrastructure buyers for the first time in three months. That doesn't prove credit losses are moving upstream. The baskets have different constituents, and GPU depreciation primarily hits whoever owns the hardware. The more interesting exposure is supplier financing. A chipmaker can book revenue from selling hardware while also holding equity, loans or guarantees tied to the customer buying it. If that customer's economics deteriorate, the supplier may face both weaker future orders and losses on its financial exposure. My hypothesis: investors will increasingly distinguish AI hardware demand funded by independent customer cash flows from demand supported by capital flowing back through the supply chain. The test is disclosed financing exposure, not CDS spreads alone. The question isn't just who sells the chips. It's who finances the buyer—and who absorbs the loss if the economics fail.

The New York Times Sued Brave. The Bigger Target May Be AI's Data Supply Chain. The NYT and other publishers have sued Brave, alleging that its search index and API distribute copyrighted news content to AI companies without permission. Anthropic, Perplexity and Mistral are among the companies named in the allegations. The claims remain unproven. The interesting distinction is between two uses of web data. An AI company can use a search API to retrieve information for a live answer, or obtain content for model training. Those are different technical and legal activities. This lawsuit puts the intermediary under scrutiny, not just the model developer. My hypothesis: the competitive advantage in AI search infrastructure may gradually shift from how much of the web a company can index to how much content it can legally deliver. The test will be licensing costs, API restrictions and whether customers pay a premium for verified content rights. A large search index is valuable. A legally usable one may become more valuable.

SoftBank's $3.7B AI Infrastructure Hit Raises a Bigger Question: Who Bears the Risk? SoftBank recorded roughly $3.7B in write-downs tied to equipment purchased for US AI data centers after project plans changed. The exact cause and equipment involved remain unclear. But the economics are worth watching. Developers commit capital years before projects generate revenue. If plans change, they may absorb the loss while suppliers retain deposits or resell scarce production slots. My hypothesis: equipment scarcity may protect manufacturers more than their customers. For power-equipment suppliers, the key test is whether cancelled orders can be replaced at similar prices without sacrificing margins. A failed AI infrastructure investment doesn't necessarily mean a failed equipment order.

Cloudflare's Gross Margin Fell. Its Operating Margin Hit a Record. Workers Explains Why. Cloudflare grew revenue 39% in Q3, its fastest pace in nearly four years. But the more interesting result sits between two margins. Non-GAAP gross margin fell 290 basis points to 72.4%. Operating margin rose 220 basis points to a record 17.5%. Management identified Workers as a major reason. Workers carries a slightly lower gross margin than some of Cloudflare's established products. But it also has a significantly lower cost to book. That changes the economics of growth. A product can generate less gross profit per dollar of revenue while contributing more efficiently to operating profit if acquiring and expanding customers costs substantially less. Cloudflare's enterprise contracts reinforce the mechanism. Its pool-of-funds model lets customers expand into additional products without renegotiating every purchase. Net retention reached 122%, and large customers accounted for 75% of revenue. The company now expects GAAP profitability in Q4, well ahead of its previous 2028 deadline. The open question is whether Workers can sustain that advantage as agent workloads become more compute-intensive. Its isolate architecture may reduce execution overhead, but growing compute demand still requires infrastructure investment. My hypothesis is that Cloudflare's competitive advantage may increasingly depend on the economics of distributing and selling compute—not just the gross margin earned on each unit. The next test is whether operating margins continue expanding while capital intensity and free cash flow remain under control. Lower gross margins aren't necessarily a weakness when the cost of winning and expanding business falls even faster.

Anthropic's Suppliers May Become Its Investors. Who Ultimately Funds AI Demand? Anthropic's withdrawn IPO pricing document reportedly outlined a potential $69.3 billion offering, alongside a separate $26 billion strategic investment involving Nvidia, Microsoft, Google, AMD, Broadcom, Micron and TSMC. The pricing information was subsequently flagged as inaccurate. Neither the IPO terms nor the strategic investment should be treated as finalized. But the proposed structure raises a more interesting question. Several companies that stand to benefit from Anthropic's infrastructure spending may also help finance it. That creates two distinct sources of growth: revenue from selling AI infrastructure, and capital invested in the companies buying it. Both can be economically rational. But they are not independent. Supplier-backed financing can accelerate deployment and strengthen long-term commercial relationships. It can also make reported infrastructure demand more dependent on continued access to capital. The distinction matters as AI labs commit to enormous multiyear compute purchases. My hypothesis is that investors will increasingly need to separate demand funded by recurring customer revenue from demand supported by financing within the AI supply chain. The next evidence is not the IPO valuation. It's the final investment terms, any linked purchasing commitments, and whether Anthropic's operating cash generation can progressively fund more of its infrastructure obligations. AI infrastructure orders can be real while the financing behind them becomes the bigger investment risk.

Claude Fable 5.2 Is 20% Cheaper. The Bigger Change May Be What It Can Coordinate. Anthropic's new Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down 20% from Fable 5.1. Output generation is up to 30% faster. But the more interesting numbers are in the work it can complete. On Anthropic's FrontierCode 2.0 evaluation, Fable 5.2 scored 49.1%, up from 32.4% for Fable 5.1. At High effort, it roughly matched Opus 5.5's best result for half the cost per task. Anthropic also reports that Fable coordinated more than 1,000 subagents over eight days to build a prototype operating system. In another experiment, a team of roughly 1,000 agents formalized Fermat's Last Theorem in 30 hours. These are company-reported results, not proof that either workload can be reproduced economically in production. Still, they point to a change in how frontier intelligence may be used. Anthropic now explicitly positions Fable as an Orchestrator or Advisor, with smaller models handling simpler work. That changes the economics of model selection. A more expensive model can be the cheaper choice if its planning reduces wasted execution, failed attempts and human intervention. Conversely, a lower token price means little if an agent repeatedly calls tools, loses context or produces work that must be redone. There is also a limit to the argument. Anthropic still recommends Opus 5.5 for most everyday work. Fable's advantage is concentrated in difficult, open-ended tasks—not every request. My hypothesis is that frontier models are increasingly competing on the cost of coordinating work, not just generating answers. That could shift value toward models that make reliable decisions about delegation, verification and when to stop. The next evidence won't be another headline benchmark. It will be whether Fable-led systems complete more complex, independently verified tasks with fewer total model calls, less rework and lower end-to-end cost. The most valuable model may not be the cheapest one to call. It may be the one that prevents the most expensive mistakes.

AI Demand Is Growing. The Cash Flow Is Telling a Different Story. Google Cloud revenue grew 103% in Q3. AWS grew 39%. Yet Alphabet spent $51.4 billion on capital expenditures in a single quarter, while Amazon expects roughly $220 billion in cash capex for 2026. Demand is not the obvious weak link. The more interesting question is how much capital it takes to serve that demand—and when that capital starts earning an adequate return. Amazon generated negative $42.4 billion in free cash flow over the trailing twelve months, even as AWS operating income increased. Alphabet reported negative $9.1 billion in quarterly free cash flow, partly affected by TPU inventory and a regulatory fine. These are not identical measures, and neither proves that AI investments are failing. But they expose a tension the market can no longer ignore. Cloud revenue can accelerate while cash flow deteriorates. New infrastructure consumes cash before it produces revenue. Depreciation then reaches earnings over several years, while power, networking and operating costs continue. The risk isn't simply that AI spending slows. It's that the capital required to sustain growth may rise faster than the cash returns investors expected. This matters differently across the semiconductor supply chain. For memory and equipment suppliers, hyperscaler spending is revenue opportunity. For hyperscalers, the same spending is a capital obligation that must eventually earn its cost of capital. Strong demand can therefore support chip orders while putting pressure on the economics of the companies placing those orders. My hypothesis is that investors are beginning to price this distinction more aggressively. The next evidence isn't another AI revenue growth figure. It's incremental returns on invested capital, free cash flow after infrastructure spending, depreciation growth and the timing of new capacity entering service. AI demand can remain strong while the economics of supplying it become less attractive. That is a very different risk from an AI demand collapse.

AI Models Are Getting Cheaper. Why Are Agents Still So Expensive to Run? Anthropic cut the price of its high-end models. Meta pushed the price of repetitive agent execution down to $0.10 per million input tokens. Yet Grok Bot has been dealing with the opposite problem: usage allowances exhausted too quickly, questions about model quality, and serving costs that Elon Musk has acknowledged are difficult to sustain. On November 10, Grok Bot reset its weekly usage limits. These developments point to a contradiction in the emerging agent economy. The price of intelligence is falling. The cost of delivering reliable work may not be falling nearly as fast. A chatbot typically answers one request. An autonomous agent may make hundreds of model calls, pass context between them, invoke tools, retry failed steps and verify results. That creates three different cost problems. The first is inference. Anthropic and Meta are attacking it from opposite directions: cheaper high-end reasoning and cheaper routine execution. The second is coordination. Dividing work among models saves money only if the overhead of routing, context transfer and retries doesn't consume the savings. The third is verification. A low-cost model can produce an answer that looks complete but is wrong. If the system cannot detect that failure, its reported cost per task understates the true cost of useful work. Grok Bot adds a commercial dimension. A generous usage allowance can attract subscribers. But if heavy users consume substantially more inference than their subscriptions cover, greater engagement can weaken rather than improve unit economics. Resetting allowances may address immediate user frustration. It does not, by itself, solve the underlying cost problem. This is where the competitive question changes. For much of the model race, companies competed on intelligence, benchmarks and token prices. Agent platforms increasingly have to compete on something harder: how much reliable work they can deliver for each dollar spent. That distinction could reshape where profits accumulate. Model providers may face lower prices for routine inference even as total usage rises. Agent platforms that manage routing, verification and recovery could capture more value, but only if those capabilities reduce the full cost of execution. My hypothesis is that cheaper inference will not eliminate the economics problem. It will expose the next bottleneck: deciding when an agent's work is actually good enough to trust. The next evidence won't be another model benchmark or a refreshed usage meter. It will be completed tasks per dollar, verification costs, retry rates, undetected failures and serving costs per paying customer. Those numbers will tell us whether agents are becoming genuinely economical—or whether cheaper tokens are simply enabling more expensive workflows.

Investors have rewarded $NVDA for Jensen Huang's leadership. But the same concentration of power may carry a governance discount. The moat is real. So is the key-person risk. What matters for valuation is the net effect. My latest analysis.

VPG's Q3 results reveal an important distinction in the AI infrastructure trade: procurement demand is spreading, but commercial maturity varies sharply by end market. Three signals stand out. 1. Optical networking and semiconductor equipment are seeing real procurement activity. VPG's Sensors segment recorded a second consecutive quarterly booking record of $51.7M, up 8% sequentially, with a 1.35 book-to-bill ratio. Orders for precision resistors from semiconductor front-end, back-end and process-control equipment makers remained strong. Bookings from high-speed fiber-optic transmission equipment manufacturers also remained elevated. This reinforces the demand signals from Ciena and Nokia, but does not establish new orders or market-share gains for either company. Nor does it identify KLA as a VPG customer. 2. Defense demand has stronger near-term earnings visibility. VPG reported sustained demand across established and next-generation aerospace and defense programs. Curtiss-Wright's latest results provide independent confirmation: Q3 orders reached $1.1B, up 15%, with a 1.1x book-to-bill and $4.6B backlog. The connection is through overlapping instrumentation and measurement requirements, not a disclosed VPG-Curtiss-Wright supply relationship. 3. Humanoid robotics remains much earlier in commercialization. VPG recorded only about $200K in humanoid bookings and $300K in revenue. Its initial customer pushed the production ramp to Q4, with orders expected to grow gradually before accelerating through 2027. That is a meaningful timing warning for investors pricing in near-term humanoid supply-chain revenue. There is also a broader quality-of-growth issue. VPG's adjusted orders exceeded $100M, but adjusted operating margin was 4.9%, down from 6.9% a year earlier. Strong procurement demand does not automatically translate into stronger earnings. My takeaway: AI infrastructure demand is broadening into optical transmission and semiconductor equipment, while defense provides a separate, more established demand channel. Humanoid robotics remains a future opportunity, not yet a material earnings driver. The distinction between orders, production ramps and profitable revenue matters more than the number of industries adopting AI. Sources: $VPG Q3 FY2026 earnings release and November 4 call; Ciena Q3 FY2026 results; Nokia Q3 2026 interim report; KLA Q1 FY2027 results; Curtiss-Wright Q3 2026 results.

Terafab's biggest risk may change before the first wafer ships. The reported target is extraordinary: 1 million wafers per month after 2032, roughly 700,000 for memory and 300,000 for logic. But those two production lines face different constraints. Leading-edge logic depends on process transfer, equipment availability and yield learning. Memory brings its own manufacturing and qualification challenges. Neither becomes production capacity simply because a cleanroom is completed. Then comes a different problem. At 1 million wafers a month, Terafab would need enough sustained demand to justify an enormous installed manufacturing base. Tesla's vehicles and Optimus, SpaceX's orbital computing plans, and xAI's infrastructure could create that demand. But their future chip requirements are not yet independently verifiable at the proposed scale. There's also a capital constraint. Equipment must be ordered and funded years before the resulting wafers generate revenue. My view: Terafab faces three tests, in sequence and sometimes simultaneously. Can it secure the tools? Can it manufacture competitive chips at acceptable yields? And can its owners consume enough of those chips at an attractive cost? The most useful near-term evidence won't be another capacity target. It will be disclosed equipment commitments, financing arrangements and measurable process milestones. A million wafers per month is a manufacturing ambition. It becomes an economic plan only when supply, capital and demand converge. Source: Nikkei reporting, November 9, 2026; Tesla FY2026 Q1 earnings call; SpaceX public filings. $SPCX

NVIDIA expects roughly 70% revenue growth in FY2028. But the more interesting number may be its gross margin. On its August 26 earnings call, CFO Colette Kress gave a preliminary FY2028 growth outlook of approximately 70%, while warning that supply could remain constrained through the fiscal year. That's extraordinary demand. But scarcity cuts both ways. NVIDIA has considerable pricing power in AI computing. Yet when HBM and other memory components become scarce, suppliers gain leverage over NVIDIA too. Management acknowledged that memory price increases had exceeded expectations. The margin outlook tells the story: Q2 FY2027: 75% Q4 FY2027 forecast: 71–72% FY2028 forecast: 72–73% NVIDIA expects price increases to help restore margins. But even with revenue growing rapidly, it may not immediately recover the profitability of its current product mix. Here's the investment question. As AI infrastructure spending expands, how much of the incremental value stays with NVIDIA, and how much moves upstream to memory and other critical-component suppliers? Revenue growth measures how much demand NVIDIA can capture. Gross margin tells us how the economics of that demand are being divided. The AI boom can make NVIDIA bigger while simultaneously strengthening the pricing power of its suppliers. Source: NVIDIA Q2 FY2027 Earnings Call, August 26, 2026 — CFO Colette Kress. $NVDA $MU $TSM

The semiconductor selloff is no longer just a weekly correction. By the November 5 lows, Micron was 48% below its all-time high. Broadcom was down 41%, AMD and SK hynix 39%, TSMC 28%, and NVIDIA 26%. The damage was even greater in equipment: Applied Materials was down 57% and Lam Research 58%. These are very different businesses. Yet the market is repricing all of them. I think the common question is changing. For much of the AI buildout, investors focused on who would capture the most spending. Now they're asking which earnings can survive a slower pace of investment. Memory is particularly exposed to this shift. HBM shortages support today's pricing, but high margins also encourage capacity expansion. The market can discount a future pricing cycle before current orders weaken. Equipment faces a different problem. Chip demand can remain strong while manufacturers postpone new fabs. Equipment revenue depends on the timing of capacity additions, not simply on how many chips customers need. NVIDIA and TSMC have another advantage: their economics extend beyond a single memory pricing cycle. NVIDIA captures value through its computing platform; TSMC supplies competing chip architectures through advanced manufacturing. That doesn't make either immune to weaker AI spending. It changes how the risk reaches their earnings. Broadcom and AMD add another dimension: customer concentration, custom silicon adoption and competitive share shifts. My hypothesis is that the market is no longer pricing AI demand alone. It's repricing the durability of each company's earnings if AI capex growth slows. The next evidence won't be the stock charts. It will be memory pricing, qualified capacity additions, equipment orders, accelerator margins and hyperscaler deployment schedules. Those will tell us whether this was mainly a valuation reset—or the beginning of a deeper earnings revision.

NVIDIA's options market is setting up for a potentially volatile earnings week. $NVDA reports on November 17. A major options expiry follows just three days later. The positioning is worth watching. At the November 6 close, NVDA had approximately $73.6M in negative net gamma exposure per 1% stock move, according to Bullflow's model. The largest negative gamma concentration was at $185, close to the $186.77 spot price. The largest positive concentration was at $200. More importantly, the November 20 expiry carried roughly $40.7M in negative gamma exposure per 1% move. That structure matters. When dealers are short gamma, their hedging can reinforce price moves: selling into declines and buying into rallies. NVIDIA's options structure therefore has the potential to amplify volatility around earnings, even without a corresponding change in the company's long-term outlook. And November 20 could change that structure as expiring positions roll off or migrate into later contracts. This creates two separate questions for AI investors. First: Is the market unwinding crowded positioning? The November 17 earnings release, November 20 expiry and subsequent options positioning will help answer that. Second: Are the economics of AI infrastructure actually weakening? That requires different evidence: hyperscaler capex commitments, GPU deployment schedules, rental pricing and cash returns. A positioning reset can finish without a fundamental deterioration in AI demand. Equally, strong AI spending can coexist with declining returns on the capital deployed. The distinction is particularly important when the same stocks are being repriced for both reasons. NVDA's options structure has the potential to amplify earnings-week volatility. What happens after the expiry may tell us more about the durability of the selloff than the initial price reaction.

CoreWeave's biggest risk may not be weak AI demand. It may be the timing of its cash returns. $CRWV reported $2.58B in Q2 revenue, up 112%, with roughly $104B in revenue backlog. But it also carried approximately $35.6B in debt principal, while quarterly net interest expense reached $640M. The backlog supports demand visibility. It doesn't eliminate the cost of delivering that capacity. Three clocks matter: how long customer contracts generate revenue, how long GPUs remain economically competitive, and when the debt must be serviced and repaid. If GPU rental economics weaken before financing obligations roll off, high utilization alone may not protect shareholder returns. The next earnings report needs to show more than revenue growth. Watch capacity delivery, interest costs, capital spending and the cash returns on newly deployed infrastructure. AI demand can be enormous without every dollar invested in serving it earning an attractive return.

Samsung Electronics and SK hynix rebounded on November 9, but foreign investors were still selling both stocks. By late morning, foreign net selling reached roughly ₩145B in Samsung and ₩424B in SK hynix, while domestic institutions were buying. That looks more like a contest between bargain hunters and investors reducing risk than a confirmed turn in the memory cycle. The distinction matters. A selloff driven by AI spending fears can reverse before earnings estimates change. But a durable recovery in memory stocks needs more than improved sentiment. Watch HBM delivery schedules, customer volume commitments and contract pricing. If those hold, the market may have repriced risk faster than the earnings outlook deteriorated. If they weaken, cheaper valuations alone won't protect investors.

The semiconductor trade is no longer just about which segment benefits from AI. It's about how much of that benefit is already in the stock price. Compute has relatively strong demand visibility. Memory offers powerful earnings leverage when supply is tight, but its profits remain sensitive to pricing and contract terms. Semiconductor equipment has a different equation: future fab spending, free cash flow and shareholder returns. Networking may have equally strong demand, yet still offer weaker returns if valuations already assume years of exceptional growth. And CPUs shouldn't be treated as one trade. AI can increase CPU workloads while shifting market share between architectures and suppliers. The distinction matters after a sector selloff. A stock can become cheaper without its earnings outlook improving. Another can deliver strong earnings and still disappoint because expectations were too high. For semiconductor investors, the next decision isn't simply whether AI spending continues. It's which companies can turn that spending into durable per-share earnings at a price that leaves room for error.