Straggler Liu | AI & Semis (@StragglerLiu) header
Straggler Liu | AI & Semis (@StragglerLiu) profile picture

Tracking where AI capex hits physical constraints — and who captures the scarcity rent. Compute • Memory • Semiconductors • Power • Infrastructure

Joined January 2026
Photo by Straggler Liu | AI & Semis (@StragglerLiu)Photo by Straggler Liu | AI & Semis (@StragglerLiu)Photo by Straggler Liu | AI & Semis (@StragglerLiu)Photo by Straggler Liu | AI & Semis (@StragglerLiu)Photo by Straggler Liu | AI & Semis (@StragglerLiu)Photo by Straggler Liu | AI & Semis (@StragglerLiu)Photo by Straggler Liu | AI & Semis (@StragglerLiu)Photo by Straggler Liu | AI & Semis (@StragglerLiu)Photo by Straggler Liu | AI & Semis (@StragglerLiu)Photo by Straggler Liu | AI & Semis (@StragglerLiu)
NAND's 2027 Risk Is Supply Elasticity, Not Just Weaker Demand The memory cycle is beginning to split. TrendForce expects DRAM supply to remain constrained in 2027, while NAND could move toward a more balanced market in the second half as higher-layer transitions and new capacity lift bit output. The warning is already visible in client SSDs. PC manufacturers are reducing SSD capacities and cutting procurement as higher component costs pressure end-market demand. TrendForce expects client SSD contract price increases to slow to less than 5% in Q4 2026. But weaker consumer demand alone won't end the NAND cycle. Enterprise SSD demand remains supported by AI infrastructure spending. The real test is whether suppliers can increase bit output faster than enterprise demand absorbs it. NAND's next downturn could begin while AI storage demand is still growing — if manufacturing productivity and capacity expansion outrun that growth. Watch enterprise SSD pricing, supplier bit output and inventory. Those will matter more than headline AI storage demand.
1
1
66
TSMC's Price Hikes May Change Which Chips Get Built Taiwan's Commercial Times reports that chip designers have received foundry price increase notices ranging from 10% to over 20%. Some expect a 4–5 percentage-point gross-margin hit if selling prices stay unchanged. The figures are not a confirmed across-the-board TSMC increase. Reuters previously reported planned 2027 wafer price hikes of up to 10%, while separate reports pointed to larger increases in advanced packaging. The margin arithmetic matters. If wafer costs equal 40% of revenue, a 10% foundry price increase cuts gross margin by four percentage points unless passed through to customers. That creates a selection problem. Designers serving price-sensitive markets may struggle to recover costs. Custom ASIC suppliers with stronger differentiation have more room to negotiate, although turnkey manufacturing exposure can still pressure reported margins. The result may be more than a transfer of profits from designers to foundries. Higher wafer prices could accelerate the shift of scarce manufacturing capacity toward chips whose customers can justify paying more. The next test is not TSMC's announced pricing. It's which designers secure customer repricing before higher wafer costs reach their income statements.
31
AI's Next Scarcity Problem: Who Gets Paid Before the Debt Comes Due? A slower AI buildout doesn't necessarily make existing data centers more valuable. It makes the timing of cash flows more important. The credit market already shows the distinction. Investment-grade data center bonds traded roughly 85 bps wider than duration-matched hyperscaler debt in May. That gap reached nearly 180 bps in September and remained around 146 bps in late October. Meanwhile, consensus estimates put five major hyperscalers' combined 2027 capex at $1.17T, roughly 115% of operating cash flow. These are forecasts, not committed expenditures. Now compare Blackstone's new AWS transaction: $5B for 647 MW of powered shells, a 7% initial cap rate, 12-year leases and 2.75% annual rent escalators, backed by an AA-rated tenant. The crucial asset isn't simply the power connection. It's the ability to turn that connection into contracted cash flow before financing costs accumulate. A completed, leased facility can collect rent while delayed competitors wait for grid access. An unfinished project may face the opposite economics: interest and capital commitments running ahead of revenue. Scarcity benefits both assets in theory. Only one is already positioned to monetize it. My thesis: the next valuation divide in AI infrastructure will be determined less by announced megawatts than by the gap between debt-service dates and revenue-start dates. Watch contracted rent commencement, construction delays and refinancing terms. Those will tell us who actually captures the scarcity premium.
4
1
228
Qwen4-Max: The Real Battle Is Cost per Completed Task Alibaba's Qwen4-Max makes a more important move than its 2.4T parameter headline suggests. It keeps the same total parameter scale as Qwen3.8-Max but cuts active parameters per token from roughly 95B to 71B. On Alibaba's Work360 benchmark, it uses 35% fewer tokens at high reasoning effort. API prices also fall 25%, to $1.50 per million input tokens and $4.50 per million output tokens. Alibaba reports average task cost falling from $4.43 to $1.67, a 62% reduction. But the token and pricing changes alone imply roughly 51% lower cost. The remaining difference needs a transparent breakdown. The bigger shift is in what AI providers are competing on. For agents, the economic unit is no longer the token. It is the successfully completed task, including tool calls, retries and verification. This could expand inference demand even as the cost of each task falls: cheaper execution makes more workflows economically viable. But Qwen4-Max is not yet a universal replacement. Independent comparisons show meaningful weaknesses in terminal coding and some visual tasks, while its strongest results are in professional and document-heavy workflows. My thesis: The next phase of AI competition will be decided by how much verified work each dollar of compute can deliver—not by parameter count or token price alone. That is also the metric investors should use when assessing the next wave of inference infrastructure.
1
64
Starship V4: The Real Upgrade Is Orbital Throughput Musk's V4 targets are substantial: a ~70m upper stage, 40% longer, carrying over 60% more propellant, with nine engines instead of six. But the more important number is 200 tonnes. SpaceX's June 2026 prospectus puts V3's target LEO payload at 100 tonnes, with future generations potentially reaching 200 tonnes, possibly as early as V4. That would double target payload capacity. It does not mean propellant growth translates directly into payload growth. The engineering challenge is preserving enough propellant for recovery while accommodating a heavier payload and a larger upper stage. More vacuum engines help propulsion, but add hardware and integration complexity. The economics hinge on something bigger than payload: reusable throughput. Flight 15 demonstrated the first orbital Ship catch, but the booster was not recovered. Musk still expects weeks of refurbishment for Ship 42, not aircraft-like turnaround. My view: V4 is less about building a bigger rocket than making orbital transport scalable. The decisive test is tonnes delivered per year, at what cost, after accounting for recovery and refurbishment—not the maximum payload on a specification sheet. The 200-tonne figure remains a target, not demonstrated performance. $SPCX
40
NVIDIA Is Trimming Some Stakes, Not Its Role as AI's Financial Backstop NVIDIA's 13F, filed November 16, shows holdings as of September 30, 2026. Since June 30: • Intel: 214.78M → 161.08M shares (−25%) • Nokia: 166.39M → 72.86M (−56%) • SpaceX: 122.76M → 130.47M (+6.3%) CoreWeave (47.21M), Coherent (7.79M) and Nebius (1.19M) were unchanged. Now compare that equity reshuffling with NVIDIA's contractual commitments. Its July 26 10-Q disclosed $36B in AI-cloud service commitments and $25B in equity investment commitments. In August, NVIDIA signed a phased guarantee capped at $105B for an Ohio data center campus serving OpenAI. These are different exposures, measured at different dates—not amounts already spent or directly additive liabilities. The $36B cloud arrangement is particularly revealing. Partners buy NVIDIA infrastructure; NVIDIA commits to purchase cloud services, but partners can replace that demand with better-paying third-party customers. NVIDIA isn't merely financing AI supply. In selected deals, it is helping underwrite demand for the capacity its chips create. That supports deployment, but it also makes GPU rental economics relevant to NVIDIA's own contractual exposure. The next test isn't another 13F. It's how much of that cloud capacity clears with independent customers—and how much ultimately falls back on NVIDIA. Sources: SEC 13F; NVIDIA FY2027 Q2 10-Q. $NVDA
57
NVIDIA's $105B Guarantee Is Not a Blank Check NVIDIA's 10-Q describes a guarantee of up to $105B supporting approximately 4.25 GW of data center capacity in Ohio, intended for OpenAI under 20-year leases. But the structure matters more than the headline. The guarantee becomes effective in phases as nine facilities enter service. It covers defined portions of lease and power obligations, not the entire project cost. Exposure declines as OpenAI makes payments. NVIDIA is supporting infrastructure deployment without assuming every layer of project risk. The investment question is what happens when other AI developers cannot secure comparable credit support. A GPU order is not a financed data center. And a financing commitment is not yet operating cash flow. Watch which projects receive binding guarantees, on what terms, and whether lenders demand stronger customer commitments. $NVDA
2
1
284
Southeast Asia's AI Data Center Boom: The Risk Is in the Next Megawatt Washington's new chip controls may not stop Southeast Asia's data center expansion. They could change which projects get built. The region accounted for roughly half of Asia-Pacific's 4.8 GW of data center capacity under construction in H1 2026. Johor alone had 1,110 MW operating, 602 MW under construction and another 2,486 MW planned. The customer data complicates the bearish case. DayOne's SEC prospectus shows that its largest customer, a global short-video platform company, generated 69.2% of H1 2026 revenue. Yet customer concentration is not the same as exposure to restricted AI training. DayOne also reported 2.28 GW of legally binding customer bookings as of September 20, with an estimated $11.4B still required to complete those commitments. BIS's new rule adds licensing requirements for advanced chips shipped to Malaysia, Singapore, Thailand and Indonesia, with a conditional carve-out for qualifying US-headquartered recipients. Separately, it restricts certain large-scale remote compute access for customers linked to controlled jurisdictions, while providing an exception for qualifying commercial recommender systems. That creates two different risks. For contracted data center capacity, the questions are customer compliance, delivery and financing. For speculative GPU projects, the questions are whether new customers can replace affected demand—and at what rental price. If marginal GPU rents fall, projects without binding offtake could lose their investment case before construction begins. Existing long-term leases may remain intact. My thesis: the first measurable impact may be a widening gap between announced capacity and projects reaching final investment decision—not an immediate collapse in occupied megawatts. Watch new GPU rental contracts, customer precommitments and construction starts. Those will tell us whether the rule is reshaping Southeast Asia's buildout or merely its customer mix.
1
2
275
Blackstone's $5B AWS Deal: The 7% Cap Rate Others Can't Copy Blackstone's $5B AWS deal prices 647 MW of powered data center shells at a 7% initial cap rate. The more important question is who can finance AI infrastructure on similar terms. AWS leases all 13 buildings for 12 years, with 2.75% annual rent increases and an Amazon parent guarantee. Blackstone's vehicle acquires 75.1%, backed by roughly $1.9B in cash and $2.5B in property-level debt. AWS retains 24.9%. The portfolio generates approximately $350M in first-year base rent, or $4.9B cumulatively over 12 years. AWS unlocks capital while retaining operational use, but assumes long-term rent obligations. Whether this creates value depends on returns from redeploying that capital. Blackstone's expected initial AFFO yield is only in the low-to-mid 5% range. A 7% property cap rate is not a 7% shareholder return. Now consider the valuation risk. At $350M in annual rent, moving from a 7% to an 8% cap rate cuts property value from $5B to $4.375B. With $2.5B of debt unchanged, residual equity value falls 25%. No tenant default. No idle GPUs. Just a higher required return. But this deal has an AA-rated parent guarantee, completed powered shells and long-term contracted rent. GPU cloud operators with shorter contracts, construction exposure or weaker credit cannot assume the same financing terms. The next divide in AI infrastructure may be between cash flows that can be financed like real estate and those still priced as technology risk. Watch whether operators without hyperscaler guarantees can secure comparable long-term financing.
2
2
255
Vulcan's Failure Starts a Second Clock for Amazon Leo ULA's Vulcan failed about 10 minutes after launch on November 15, leaving 40 Amazon Leo satellites short of orbit. The cause remains under investigation. Assigning blame to a particular engine or stage now would be premature. The bigger issue is the launch schedule. Amazon contracted 38 Vulcan missions for Leo. This was the first. Before the failure, Amazon had deployed 398 satellites toward a planned 3,236-satellite constellation. Losing 40 satellites hurts. Losing months of launch capacity could hurt more. Every additional delay pushes back deployment while engineering, manufacturing and other constellation costs continue. The critical question is no longer just what failed. It's how quickly ULA can identify the cause, qualify a corrective action and return Vulcan to a dependable launch cadence. For Amazon Leo, the investigation clock is now part of the business plan.
1
80
AI Data Center Debt: The Loan Is Signed. The Syndication Is Delayed. Bloomberg reports that syndication of a $488M loan for an Actis-backed Malaysian data center serving ByteDance has been delayed. The distinction matters. HSBC, AIIB, SMBC and Bank of China have already signed the loan and committed funding. What slipped was a November 2 presentation to prospective lenders and a planned site visit. One source attributed the delay to management scheduling. Both are being rearranged. This is not a canceled loan or evidence that lenders have rejected the project. But it highlights a part of AI infrastructure financing that deserves more attention. Arranging a loan and distributing its exposure across banks are different steps. As data center borrowing grows, the ability to syndicate committed debt efficiently will matter for banks' balance-sheet capacity and the financing of subsequent projects. One delayed syndication proves little about credit appetite. Repeated delays, wider spreads or larger retained exposures would tell a different story. The next signal is whether the syndicate fills on its original terms—not simply whether the loan was signed.
72
AI Optics: The Next Fight Is Over Manufacturing Capacity Marvell's planned $115B acquisition of Lumentum says something important about where AI infrastructure is heading. Electrical and optical connectivity are becoming too closely linked to optimize separately. Meanwhile, the demand is already showing up in reported numbers. Corning's optical communications revenue reached $2.38B in Q3, up 44% YoY. FormFactor reported $280.5M in revenue, up 38%, with its systems business growing more than 20% sequentially on stronger co-packaged optics demand. Teradyne's revenue rose 70% to $1.31B, although that growth spans several semiconductor test markets. The investment question is changing. As optical engines move closer to GPUs and switch ASICs, the industry needs more than faster transceivers. It needs qualified laser sources, advanced packaging, reliable fiber connections and optical testing before expensive chips are assembled. Those capabilities don't scale at the same speed. A hyperscaler can increase its optics budget overnight. It cannot create qualified InP capacity or a mature wafer-level test process overnight. That's where I expect the next supply constraints to emerge. The companies with the strongest pricing power may not be those shipping the most optical modules, but those controlling the hardest manufacturing steps to replicate. The evidence to watch is lead times, qualified capacity and gross margins—not just another record order book.
1
93
Google TPU Is Becoming a Strategic Alternative to Nvidia Google's TPU order visibility is extending well beyond the current AI investment cycle. That says something about its competitive position. Broadcom's SEC filing confirms roughly 3.5 GW of additional TPU capacity for Anthropic starting in 2027, with its Google development and supply agreement extending through 2031. Broadcom has since outlined another 10 GW of expected Anthropic TPU deployment in 2028. These are not small experimental allocations. Customers are planning infrastructure around Google's architecture years in advance. JPMorgan's latest analysis estimates that Nvidia GPUs account for 39% of future frontier-lab compute agreements, compared with 16% for Google TPUs and 24% for Broadcom custom silicon. The latter is a separate category, not additional Google TPU share. Nvidia remains formidable. Its software ecosystem, networking and deployment scale won't disappear because customers sign ASIC contracts. But customers increasingly have credible alternatives. And Nvidia's leadership uncertainty gives large AI buyers another reason to avoid relying too heavily on a single supplier. The investment implication is broader than TPU shipment growth. Google is gaining a stronger position in the economics of AI compute, while MediaTek and Broadcom stand to benefit from a longer, more diversified custom-silicon investment cycle. Nvidia doesn't need to lose its lead for that shift to matter. It only needs to lose some of its pricing power. $GOOG $NVDA
72
The Fed Just Put a Number on AI's Financing Risk The Federal Reserve's November Financial Stability Report estimates that advanced-computing-related debt issuance exceeded $680B through October 2026, up from $130B in all of 2023. Outstanding related public debt now exceeds $1.1T. The financing mix is changing too: corporate bonds are increasingly joined by private credit, structured debt and project finance. This matters because AI infrastructure risk no longer sits entirely on the balance sheets of the companies building it. It is spreading through lenders, leases, guarantees and project structures. My concern isn't that AI demand disappears. It's that financing costs can rise faster than compute revenue, forcing otherwise viable projects to refinance, delay deployment or accept lower returns. The next AI infrastructure bottleneck may be the cost of capital—not the availability of GPUs.
1
2
317
Ship 42 Was Caught. The Refurbishment Clock Starts Now. SpaceX's Flight 15 delivered a major milestone: Ship 42 returned from orbit and was caught at Starbase roughly 22.5 hours after launch. The booster was lost, so the planned double catch remains unfinished. Musk has now inspected the ship's heat shield and engines. Both look good, he says, and Ship 42 could potentially fly again after a few weeks of refurbishment. But he also acknowledged that airplane-like reuse will require many smaller fixes to eliminate refurbishment beyond scheduled maintenance. That's the distinction investors should watch. Recovery proves a vehicle can return. Reflight demonstrates reuse. Repeated flights with minimal refurbishment are what change launch economics. The next milestone isn't another catch. It's how much work—and how much time—Ship 42 needs before flying again.
2
3
262
AI Agents May Have a Second Scaling Law A smaller model just beat a much larger one on a long-context task. Not by training a better model, but by changing how it handles information. In Recursive Language Models research, GPT-5-mini with recursive execution outperformed GPT-5 by more than 34 percentage points on a ~132K-token OOLONG benchmark. Another result is equally interesting. At ICML 2026, Context Folding researchers reported 62.0% on BrowseComp-Plus using a 32K active context. A comparable summarization-based agent scored 52.7%. The conventional RL-trained agent scored 44.6%. Same underlying model. Different execution architecture. The implication goes beyond context compression. We may be entering a period where agent performance scales along two partly independent dimensions: model intelligence and the efficiency with which an execution system turns that intelligence into completed tasks. That could change where value accumulates in AI. A company that improves verified task completion without paying for a larger model may capture economic value that previously flowed to model providers. But there's a catch. Prime Intellect's RLM experiments also found worse performance on some math tasks and substantially longer execution times. Context Folding's smaller active window doesn't establish lower total inference cost. The next competitive metric shouldn't be context length or tokens saved. It should be verified tasks completed per dollar, including orchestration, retries and latency. My hypothesis: the next major improvement in agent economics may come from execution architecture rather than another increase in model size. The benchmark evidence is emerging. The production economics remain unproven.
45
Anthropic's IPO May Reveal Where the AI Trade Is Moving Goldman's latest prime brokerage data puts hedge funds' US technology long/short ratio at 1.50, its lowest since records began in 2016. Last week, ETF short positions fell 8.6%, even as funds added single-stock shorts. That's an unusual combination: less broad-market hedging, but more selective bets against individual companies. Now Anthropic is preparing an IPO of 825 million shares at an indicated $70–78 per share, alongside a separate $26B strategic private placement. The timing matters. Investors are reducing exposure to parts of the existing technology market just as one of the largest AI equity offerings approaches. That doesn't necessarily mean money is leaving AI. It may mean investors are becoming more selective about which part of the AI value chain deserves their capital. Anthropic's listing could provide a new valuation reference for frontier models, against which investors can reassess the economics of GPU suppliers, neoclouds and AI software companies. But strong IPO demand alone wouldn't prove that rotation. The test is what happens after pricing: whether existing AI stocks lose capital to the new listing, or whether a successful debut lifts valuations across the sector. The AI trade may be changing ownership before it changes direction.
57
OpenAI vs. Anthropic: Revenue Is Growing. But Who Is Converting It Into Profit? OpenAI generated approximately $10B in Q3 revenue, up 50% sequentially. Yet its adjusted operating loss widened to $10.9B from $10.1B in Q2. Its adjusted operating margin improved from -150% to -109%, but the business is still losing more than $1 for every dollar of revenue. Anthropic reported $16.96B in Q3 revenue, up 47%. Its GAAP operating loss was $567M, but its adjusted operating income reached $2.46B, a 15% margin. The comparison needs one important qualification: Anthropic's adjusted result excludes $1.07B in stock compensation and $1.96B in non-cash charitable charges. OpenAI's reported adjusted loss also excludes $2.6B in stock compensation. These are not fully comparable accounting measures. The deeper difference is in compute economics. Anthropic spent $9.86B on compute and infrastructure in Q3, including $5B under expensive short-term commitments. This is not simply a contest of model capability or revenue growth. It is a test of how efficiently each company converts compute spending into recurring enterprise revenue and sustainable operating margins. OpenAI is scaling revenue rapidly. Anthropic is currently demonstrating better adjusted operating economics. Neither has established that frontier AI can sustain its growth without enormous continuing investment. The next question for investors is not who grows faster, but who can turn each additional dollar of compute into more durable profit.
48
Sovereign AI Is Creating a New Dependency on the Global Chip Supply Chain Governments are spending billions to gain control over AI. But the more they build, the more they depend on the companies supplying the technology. CNAS tracked 184 government-backed sovereign AI projects across 67 countries and the EU by June 2026. More than 60% disclosed at least one foreign partner, and four-fifths of those partners involved US companies. NVIDIA GPUs appeared in 45% of tracked infrastructure projects. The national strategies differ, but the underlying constraint is remarkably similar. France's Mistral trained its 1-trillion-parameter Large 4 model on 3,800 NVIDIA Grace Blackwell GPUs in European data centers. Its €21B-plus valuation and €3B funding round demonstrate that Europe can build competitive AI companies. They do not demonstrate independence from foreign compute. Saudi Arabia's HUMAIN has expanded its NVIDIA partnership to 84MW and approximately 27,000 GPUs. The first 14MW phase covers around 6,000 Blackwell GPUs; the remaining capacity is planned for later deployment. Saudi Arabia can own the infrastructure and control its local deployment, but the underlying accelerator technology remains externally supplied. India is taking a different route. Sarvam's 30B and 105B models and BharatGen's multilingual models target local languages, public services and national data requirements. This creates meaningful application and data sovereignty without requiring India to manufacture the world's most advanced accelerators. The deeper issue is that AI sovereignty is not a single technology problem. Advanced accelerators depend on foundries, lithography, HBM, advanced packaging, networking, power equipment and specialized talent. Controlling a data center does not mean controlling the supply chain that keeps it running. The economics reveal a paradox: every new sovereign AI project expands demand for the same concentrated global suppliers, potentially deepening dependence even as governments seek independence. The realistic strategy is not full technological autarky. It is selective control: secure domestic data and deployment, diversify critical hardware suppliers, build capabilities in competitive niches, and preserve access to global technology. The next contest in AI may not be about which country builds the most sovereign data centers. It may be about which country can secure the most strategic position in a supply chain that nobody can fully own.
1
2
98
Retail Can Buy NVIDIA. It Cannot Finance the AI Credit Cycle. Nearly $2B in retail money flowed into NVIDIA shares during the week Jensen Huang resigned, according to Vanda Research data reported by the WSJ. That can support the stock price. It cannot refinance a neocloud's debt. JPMorgan's November 14 analysis exposes a more consequential risk: NVIDIA has become more than a chip supplier. Over the last three quarters, it deployed 58% of operating cash flow into equity investments and acquisitions. Related off-balance-sheet debt and lease liabilities across its ecosystem rose from $214B to $422B, or $530B including certain infrastructure guarantees. Meanwhile, AI infrastructure high-yield debt now yields around 10%, against roughly 8.3% for the broader high-yield market. This is where Jensen's departure becomes financially significant. If NVIDIA changes its approach to customer financing, revenue guarantees and infrastructure commitments, some neocloud operators may face higher borrowing costs precisely when they need more capital to expand. November 17 earnings will test demand, margins and guidance. But the bigger question is whether NVIDIA's next leadership team will continue supporting the financing structure behind its customers' growth. A retail buying surge can cushion an equity selloff. It cannot replace vendor financing, bank credit or sustainable project cash flow. The real stress test is not whether investors still believe in NVIDIA. It is whether the AI infrastructure ecosystem can finance its next expansion without NVIDIA acting as its financial backstop. $NVDA
48