Straggler Liu | AI & Semis

AI data center announcements keep getting bigger. The gap between announced capacity and deliverable capacity is becoming harder to ignore. Drax, Brookfield and Radiant are targeting over 1 GW in the UK, backed by a site with 4 GW of grid access. Yet the first 100 MW, targeted for Q1 2028, still needs planning and network approvals, a tenant and final investment decisions. IREN offers a different warning. Its fully contracted, fully funded $4B+ ARR target has shifted from year-end 2026 to January 2027, potentially February, following a customer-requested change and supplier timing. Demand hasn't disappeared. But secured power, committed capital and signed contracts don't eliminate execution risk. The metric that matters now is how much capacity clears every gate and reaches billable service on schedule.

Alchip's Q3 tells a more interesting story than 267% revenue growth. Revenue hit $819M as its 3nm AI accelerator entered volume production. Yet gross margin fell below 20%. That's the tension in the custom ASIC business. Design fees can carry attractive margins. Once a program moves into mass production, revenue can multiply while the mix shifts toward lower-margin silicon shipments. Alchip also says wafer, substrate and packaging shortages cost it roughly $100M in potential quarterly revenue. And its next 2nm accelerator is approaching tape-out, with volume production expected to contribute from late 2027. For investors, the question is no longer whether hyperscalers want custom silicon. Alchip's production ramp makes that clear. It's how much economic value the design partner retains as volumes scale. The next ASIC winners won't be determined by tape-outs alone. Watch operating profit growth, packaging allocation and the transition from one chip generation to the next. Revenue captures the scale of the opportunity. Margins and cash conversion tell us who keeps the economics.

The next AI market shock may start in credit, not earnings. Blackstone's BXDC is approaching its November IPO lock-up expiration. Separately, its private-credit fund BCRED has faced elevated redemption requests, while CoreWeave's five-year CDS has traded above 1,000 basis points. These aren't the same risk, and they don't establish a forced-selling event. But they expose a vulnerability in the AI buildout. Compute infrastructure is increasingly financed with debt against assets whose economic value depends on future utilization, rental rates and technology cycles. When credit spreads widen, the first casualty may not be AI demand. It may be the next round of financing. That creates a feedback loop: higher funding costs weaken project economics, delay deployments and put pressure on leveraged operators. Stronger borrowers can then gain access to capacity and customers on better terms. The market may be pricing AI companies as technology businesses while some of their infrastructure partners increasingly behave like leveraged asset-finance vehicles. Watch refinancing terms, collateral requirements and actual asset sales—not just stock prices or lock-up calendars. A credit squeeze doesn't need to become a systemic crisis to reshape who captures AI's profits.

Micron and AMD expose two very different ways to make money from the AI buildout. Micron's latest guidance implies roughly $51B in quarterly non-GAAP operating profit, assuming its revenue, margin and expense targets hold. That's extraordinary earning power for a memory manufacturer. But the valuation question isn't simply how much Micron earns today versus AMD. It's how much of those earnings survives the next memory supply cycle. Memory profits respond sharply to pricing, capacity additions and inventory conditions. AMD's economics depend more on product execution, accelerator adoption, software competitiveness and its ability to sustain margins against much larger rivals. One company is monetizing a severe supply constraint. The other is investing to capture a larger share of the compute market. Both can benefit from AI spending, but their earnings deserve different durability assumptions. For $MU , watch DRAM pricing, HBM supply commitments and capacity expansion. For $AMD , watch accelerator revenue, gross margins and customer adoption. The market shouldn't value a peak-cycle dollar of profit the same way it values a dollar that can compound for years.

Friday's chip rebound repaired prices, not the investment thesis. Nasdaq +1.09%. SOX roughly +4%. But the Nasdaq still lost 3.4% for the week, its fourth consecutive weekly decline. Thursday alone had erased 5.5% from the Nasdaq and roughly 12% from semiconductors. The distinction matters. A rebound after forced selling tells us buyers are returning at lower prices. It doesn't tell us whether AI infrastructure spending, data-center permitting or financing assumptions have stabilized. Those risks need separate evidence. Next week, I'd watch whether Asian semiconductor stocks hold up, whether credit conditions tighten further, and whether equipment and infrastructure companies revise their delivery expectations. If chip prices recover while financing conditions keep deteriorating, the rally deserves skepticism. If both stabilize, the case for a more durable recovery gets stronger. One green session isn't enough to settle that question.

Texas turning blue is a political earthquake. For AI investors, the bigger question is what happens to projects that take longer to build than an election cycle. A semiconductor fab, a major grid upgrade or a large AI campus cannot be planned around two-year political swings. Capital is committed years before the first meaningful returns arrive. The 2026 midterms have raised fresh questions about AI regulation and data center permitting. Some scrutiny is justified. Communities have legitimate concerns about electricity prices, water use and who pays for infrastructure. But uncertainty has a cost. If developers cannot predict permitting requirements, grid access or future operating restrictions, lenders demand more protection. Projects become harder to finance, and some never reach construction. That doesn't necessarily reduce AI demand. It can reduce the infrastructure available to serve it. For investors, the distinction matters: tougher rules may hurt speculative projects while increasing the scarcity value of facilities with secured power, permits and customers. The policy challenge isn't choosing between technological progress and accountability. It's creating rules that protect the public without making long-term investment impossible.

China's semiconductor equipment story is moving beyond import substitution. ACM Research's Q3 results offer a useful look at what comes next. Revenue rose 36.6% to $367.7M. Operating income nearly doubled to $55.5M. The company also raised its 2026 revenue outlook to $1.24B–$1.28B. The technology progress is real. ACM is expanding its cleaning applications and has shipped a 310 × 310 mm panel electroplating tool to a customer outside mainland China. But look at the balance sheet. Inventories reached $851M, up from $703M at year-end. Receivables climbed to $705M from $504M. Neither figure proves weak demand. Equipment makers need inventory to support long production cycles, and receivables can rise with shipments. But both tie up capital before cash arrives. That creates an important distinction for investors. Winning a process qualification is not the same as winning repeat production orders. Shipping equipment is not the same as customer acceptance. And reported earnings are not the same as cash generated. As Chinese equipment suppliers expand into more complex processes, their competitive advantage will increasingly depend on how efficiently they turn engineering capability into accepted, repeatable, cash-generating production. The next phase of localization may produce stronger technology companies without rewarding every shareholder equally. For $ACMR and its peers, I'd watch repeat orders, acceptance cycles, inventory turnover and operating cash flow alongside revenue growth. That's where the next valuation differences may emerge.

DeepSeek just made a useful point about coding agents. Not every task needs a frontier model thinking at full throttle. In Isoquant's Tetris demo, DeepSeek V4.1 Flash finished in roughly 10 seconds versus around 90 seconds for Claude Opus 5.5, streaming at 600+ tokens/sec. It's a vendor demo, not a controlled benchmark. DeepSeek ran without reasoning; Opus used its default setting. But the economic question matters more than the speed comparison. Once a cheaper model can reliably complete a task, additional intelligence has to justify its cost in latency, retries and verification. That changes how coding agents should be built. Use fast models for routine, verifiable work. Escalate when complexity or failure risk warrants it. The winner won't necessarily be the model with the highest benchmark score. It may be the system delivering the most accepted work per dollar and per minute.

The AI valuation reset is reaching the private market. Bloomberg reports that prospective Anthropic IPO investors are discussing valuations of $1.2T–$1.3T, with some as low as $1.1T, versus earlier expectations of $1.4T–$1.5T. OpenAI is seeing similar pressure. Its initial $11B financing came at a $1.25T valuation, below its earlier $1.4T target. The planned round has also reportedly shrunk from $30B to roughly $20B. These are different transactions, but they point to the same shift: investors are becoming more selective about what they're willing to pay for AI growth. OpenAI has already secured capital. Anthropic still has to test what public-market investors will pay. The real question isn't the headline valuation. It's how much capital these companies can raise at that price. AI demand can keep growing while the price investors pay to own that growth falls.

SpaceX reportedly plans 6,000 racks for orbital data centers in 2027, implying roughly 1 GW of compute capacity. If confirmed, this is more than a launch story. Nearly every watt consumed by the hardware eventually becomes heat. In orbit, that heat must be rejected primarily through radiation. The critical metric isn't just $/kg to orbit. It's how much reliable compute each kilogram of power and thermal infrastructure can support. Cheaper launches help. But thermal engineering may determine whether orbital AI can compete economically with terrestrial data centers.

Terafab's real ambition may be faster design-to-silicon iteration, not just more fab capacity. But faster tapeouts don't guarantee faster learning. Yield, process stability and test turnaround still set the pace. Working silicon will tell us more than capacity targets.

NVIDIA's $108B earnings test just got harder. After the midterms, AI faces a new bottleneck: data-center permits, grid costs and political scrutiny. What if GPUs can ship faster than customers can deploy them?

AMD's Q3 results point to something larger than GPU competition. Data Center revenue rose 97% to $8.55B, with record demand for both EPYC CPUs and Instinct accelerators. Agentic AI could make that combination increasingly important. A GPU handles model inference. But agents also need CPUs and system memory for tool execution, databases, orchestration and persistent application state. That creates a potential second demand engine alongside accelerator spending: server CPUs and conventional DRAM. For investors, the implications aren't equal. Memory suppliers and advanced foundries may benefit from tighter capacity, while PC manufacturers could face higher component costs. The next test is whether CPU and DDR5 demand grows faster than GPU deployments alone would explain. If it does, the AI infrastructure cycle is broadening beyond accelerators — not simply rotating from one chip vendor to another.$AMD

Micron closed at $656.39 Thursday, down 17.5% in one session and 47.7% from its 52-week high of $1,255. The important question isn't whether HBM demand disappeared overnight. There's no evidence it did. It's whether investors are starting to price the next memory cycle before the current one has peaked. HBM allocations can remain tight while expectations for future pricing, capacity growth and margins deteriorate. That distinction matters. A memory producer can report strong earnings and still see its valuation contract if the market no longer believes those earnings are sustainable. I'd watch HBM contract pricing, customer delivery commitments and conventional DRAM supply before calling this either a bargain or the start of a new downcycle. The stock has already repriced. The industry fundamentals still need to be tested.

The U.S. AI debate now has two separate regulatory fronts: where compute can be built, and how far frontier models can be developed. The infrastructure risk is measurable. SemiAnalysis estimates 33.8 GW of planned U.S. capacity in 2027, with roughly 75% already under construction. The much larger 2028 pipeline is more exposed to permitting delays. Separately, Sanders and Casar have introduced legislation that would pause certain advanced AI development until federal safety rules are established and prohibit artificial superintelligence. It is a proposal, not law. The strategic question is what happens if both constraints become binding. China could gain relative development time, but only if its available chips, interconnects and training efficiency support continued progress. The metric to watch isn't announced GW alone. It's usable compute delivered on time — and the model capability produced from it. Regulatory uncertainty matters. So does the ability to turn compute into intelligence.

Restricting AI data centers in the U.S. won't necessarily reduce global demand for compute. It may change where that compute gets built. If permitting delays become unpredictable, hyperscalers have another option: move suitable workloads and future capacity to jurisdictions where power and construction timelines are more reliable. Training clusters are more geographically flexible than latency-sensitive inference. Not every workload can move, and export controls still apply. But the investment consequence is important. A U.S. construction slowdown could become a geographic redistribution of AI CapEx rather than an equivalent reduction in spending. The winners wouldn't simply be countries with cheap electricity. They'd be operators able to deliver permitted, powered capacity on schedule. The question is whether future project announcements begin shifting across borders before U.S. demand actually slows.

The Nasdaq 100 lost roughly $2.5 trillion Thursday as investors worried that tighter data-center rules could slow the AI buildout. But lower capacity growth doesn't necessarily mean lower CapEx. New grid requirements, water restrictions and the loss of tax incentives could make each delivered megawatt more expensive, even as fewer projects get built. That's not automatically bullish for infrastructure suppliers. A hyperscaler spending more to deliver less capacity may generate higher CapEx but weaker returns. Meanwhile, operators with permitted, energized sites could gain bargaining power as new supply becomes harder to bring online. The number I'd watch isn't total AI CapEx alone. It's the cost per revenue-generating megawatt — and how much cash that megawatt ultimately earns.

AWS is growing 39%. Amazon's free cash flow is −$42B. Now comes an $80B infrastructure financing facility—and even higher CapEx expected in 2027. The question isn't demand. It's how much cash that demand will generate.

KOSPI plunged more than 8% shortly after Friday's open, triggering a market-wide circuit breaker. Samsung and SK hynix were hit hard again. This is more than Seoul following Nasdaq. Thursday's U.S. selloff repriced AI infrastructure spending and regulatory risk. Korea is now repricing the memory suppliers that sit upstream of nearly every AI accelerator. But a collapse in HBM suppliers' share prices doesn't mean HBM supply has suddenly loosened. The real test comes in customer behavior: do GPU makers maintain their HBM allocations and contract commitments, or start pushing out deliveries and renegotiating terms? That's where a financial-market shock would become a change in industry economics.

SpaceX's five-year CDS hit a record 284.3 bps Thursday, according to Bloomberg/ICE. That's notable for a company with nearly $90B in cash and short-term securities, $38B in debt and a BBB+ rating from Fitch. The pressure isn't simply today's balance sheet. SpaceX has committed to a roughly $40B GPU financing lease. Its Super Intelligence business is growing rapidly, but one customer accounted for 34.5% of consolidated Q3 revenue, and some cloud contracts allow termination on 90 days' notice. That creates a duration mismatch worth watching: multi-year equipment financing against less certain customer revenue. The real test is whether contracted compute demand turns into durable cash flow before financing commitments become a burden. Credit spreads may reveal that risk before reported revenue does.