Straggler Liu | AI & Semis

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

[Straggler Liu | AI & Semis @StragglerLiu]

Apple Is Redefining "AI Sovereignty" — Bringing Billion-Parameter Models from the Cloud Back to Your Desk

On August 25, 2026, Apple released two new products: the Mac mini with the M6, its first 2nm chip, and the Mac Studio with the M5 Ultra, its first quad-die architecture.

This is not a routine hardware update. It's Apple's definitive answer to where AI should run — locally, not in the cloud.

I. Mac mini M6: 2nm "AI Democratization"
The M6 is Apple's first 2nm chip and the first M-series chip to use three CPU core types: 2 super cores, 4 performance cores, and 6 efficiency cores. The 12-core CPU has 2 more cores than the M4, and the 12-core GPU also has 2 more cores.

The real breakthrough is in the AI architecture:

For the first time, every GPU core has a built-in Neural Accelerator. AI compute no longer depends on a separate Neural Engine — the GPU itself accelerates AI. Combined with the Dual 16-core Neural Engine (32 cores total), the M6 Mac mini processes LLM prompts in LM Studio up to 4.8x faster than M4 and 13.5x faster than M1.

32GB unified memory, 170GB/s bandwidth — the M6 Mac mini can comfortably run 7B-13B parameter models.

**Starting at $899**. For the first time, a 2nm chip appears in a $900 device.

II. Mac Studio M5 Ultra: The "VRAM Wall" Terminator
If the M6 is "AI democratization," the M5 Ultra is "AI sovereignty" in its ultimate form.

The M5 Ultra is Apple's first quad-die architecture M-series chip — connecting two dual-die M5 Max chips via UltraFusion technology into a single processor, with inter-die bandwidth exceeding 4.4TB/s. 36-core CPU (12 super cores + 24 performance cores), 80-core GPU, 512GB unified memory, 1.2TB/s bandwidth.

What does 512GB unified memory mean? The NVIDIA RTX 5090 has only 32GB of VRAM. This means the Mac Studio can run hundreds of billions or even trillion-parameter LLMs locally — Qwen3 235B MoE, DeepSeek R1 670B — models that previously required cloud or specialized servers can now run entirely on-device.

Peak AI compute performance is 4.3x that of the M3 Ultra and 9.8x that of the M1 Ultra.

Four Mac Studios can also be clustered via Thunderbolt 5, delivering up to 3x faster distributed AI inference performance than a single system.

Starting at $5,499. The 512GB memory configuration ships in late October.

III. Unified Memory vs. Discrete VRAM: Two Completely Different AI Hardware Philosophies
Apple's approach differs fundamentally from NVIDIA's:

NVIDIA: Compute chip + discrete VRAM (RTX 5090: 32GB GDDR7, 1.8TB/s bandwidth). Advantage: peak compute. Bottleneck: the VRAM wall — model parameters must fit within VRAM capacity.

Apple: Compute chip + unified memory (M5 Ultra: 512GB, 1.2TB/s bandwidth). Advantage: "memory is VRAM" — entire models can be loaded without partitioning or compression.

The unified memory architecture allows entire models to reside in memory, eliminating the need to copy data between CPU and GPU. 512GB capacity means trillion-parameter models can be fully loaded, which is critical for inference latency and throughput.

This is not about追赶 NVIDIA's peak compute — it's about redefining the boundaries of AI computing.

IV. AI Sovereignty: Apple's Strategic Bet
Apple is fighting an "AI sovereignty" war. The core logic: your data, your models, your hardware — all local, without passing through any cloud API.

This has profound implications for privacy-sensitive industries like finance, healthcare, and legal. When data cannot leave the premises, local AI deployment is not optional — it's mandatory.

As Ars Technica notes, the local AI inference boom on Macs began with the macOS 26.2 update in December 2025, which enabled low-latency Thunderbolt 5 communication for distributed AI inference using MLX. Since then, developers and researchers have been daisy-chaining multiple Macs to run models far larger than any single device could handle. This release is Apple's official acknowledgment and systematic promotion of that trend.

Apple's goal is not to become the next NVIDIA — it's to become the default hardware standard for the "AI sovereignty" era.

V. Risks and Uncertainties
Software ecosystem remains unproven: The degree of optimization for mainstream frameworks like PyTorch and TensorFlow on Apple Silicon will determine its real-world value.

32GB is still a bottleneck for the Mac mini: Models above 70B parameters still require the M5 Pro (64GB) or Mac Studio.

Significant price increases: Mac mini starting price rose from $599 to $899; M5 Ultra Mac Studio rose from $5,299 to $5,499. Apple is charging a premium for AI performance.

Performance data is Apple's internal testing: Independent benchmarks are needed for verification.

NVIDIA and AMD won't stand still: NVIDIA's Rubin and AMD's MI455X are expected to enter production in 2027 and could close the gap.

VI. Conclusion
With the dual-track strategy of the M6 Mac mini and M5 Ultra Mac Studio, Apple is building a complete local AI computing ecosystem spanning "entry-level AI development" to "frontier AI research."

The M6 brings 2nm chips and local AI capabilities to the mainstream market. The M5 Ultra makes it possible, for the first time, to run trillion-parameter models locally — without relying on cloud APIs, without data privacy concerns, without paying ongoing token fees.

As cloud AI costs continue to rise and data privacy becomes increasingly sensitive, Apple offers a third option: own your AI hardware, run your AI models locally, control your AI data.

This is not about faster chips — it's about who controls the next decade of AI. Apple is betting on one thing: users want to own their AI, not rent someone else's.

https://t.co/HfuJoOvti3

[Straggler Liu | AI & Semis @StragglerLiu]

The optics here aren't in the earnings revision. They're in the multiple compression that ate it. 4-8% EPS upgrades, 29.3x → 27.8x, target price moves 633→639. Market is saying: we believe the supply story, we're less convinced the margin story holds at scale. Eoptolink's transition from surge to scale is priced as a volume trade, not a margin compounder.

[Straggler Liu | AI & Semis @StragglerLiu]

OpenAI Jalapeño: The Customer Who Became a Competitor

OpenAI has long been NVIDIA's most important customer. On August 25, it became something else: a potential competitor.

At Hot Chips 2026, OpenAI released the first real-world benchmarks for Jalapeño, its custom inference chip. The numbers are impressive. But the story here isn't just about performance — it's about a structural shift in the AI hardware supply chain.

What Jalapeño Is

Jalapeño is an ASIC designed specifically for large language model inference, with architecture led by OpenAI and chip implementation by Broadcom. From design start to tape-out took roughly 9 months.

SemiAnalysis was invited into OpenAI's lab for independent testing and concluded that Jalapeño "beat every Nvidia, AMD, and Google chip we were able to test."

Specifications

SpecificationJalapeño RackAccelerators128 chips4-bit compute1.7 exaFLOPSHBM4 capacity27.5 TBMemory bandwidth~2 PB/sTDP700W (measured sustained ≤550W)
For context, AMD and NVIDIA's latest rack systems deliver 1.46x to 2x more compute, but Jalapeño wasn't designed for peak FLOPs — it was designed for inference efficiency.
The Numbers
OpenAI used SemiAnalysis's InferenceX benchmark across GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T, comparing against the best recorded numbers on NVIDIA GB200 NVL72 and GB300 NVL72 systems.
AI work per watt: 1.5x – 1.9x
End-to-end latency: 1.7x – 3.6x lower
Ultra-low-latency: 2.1x – 4.1x faster

SemiAnalysis' independent testing further confirmed that in Single-Token Prediction mode (no speculative decoding or multi-token prediction), Jalapeño achieved over 700 tokens/sec/user on DeepSeek R1, and approximately 1,400 tokens/sec/user on Kimi-K2.5 and GPT-OSS. Other chips' comparative numbers all used MTP optimization — meaning Jalapeño's lead was achieved under fairer or even more stringent conditions.

The Architecture

Jalapeño's real advantage is hardware-software full-stack co-design. OpenAI can simultaneously design models, products, serving software, chips, memory, networking, and systems, using real workload experience to optimize every layer.

Jalapeño is optimized for both critical inference phases:

Prefill phase: Compute-intensive, requiring high compute

Decode phase: Memory-bandwidth-intensive, requiring high HBM bandwidth

OpenAI hardware VP Richard Ho explicitly stated that Jalapeño is not over-optimized for OpenAI's specific models, but a general-purpose inference chip. As evidence, OpenAI successfully ported Doom to run on Jalapeño using just Codex prompts.

The Strategic Signal

Three implications stand out:

First, the economics of inference change. Jalapeño's per-watt and latency gains directly reduce inference cost per token. For OpenAI, this means higher gross margins on API services. For NVIDIA, it means its largest customer will buy fewer GPUs over time.

Second, the customer-competitor dynamic is unprecedented. NVIDIA can't simply cut off OpenAI — it's too important as a customer and an ecosystem partner. But OpenAI's long-term strategy is clearly to reduce dependency. This creates a structural tension that didn't exist three months ago.

Third, Jalapeño is a proof point for the broader industry. If OpenAI can build competitive inference silicon, any company with massive AI workloads — Google, Microsoft, Amazon, Meta — can replicate this path. The "model developer" is becoming a "chip developer."

What This Means for NVIDIA

NVIDIA closed at $211.47 on August 25, 2026. The market's reaction has been muted — Jalapeño is still "very small volumes" in 2026, with a meaningful ramp only in 2027. But the long-term signal is real.

NVIDIA's inference pricing power has been a core component of its valuation. If the largest inference customer can build its own silicon, the long-term pricing power story is at risk.

2027 is the critical year. NVIDIA Rubin and AMD MI455X are expected to ramp production in early 2027. If they match or exceed Jalapeño's gains, the threat may be contained. If not, the math changes.

The Bottom Line

Jalapeño isn't just a chip. It's a signal of power transfer in the AI hardware market. When the industry's largest customer starts building its own chips, the pricing logic of the entire supply chain changes.

The real significance isn't the 1.9x performance gain — it's that model developers can now own their own hardware. Once that path is validated, the rules of the AI chip market are fundamentally rewritten.

The next 12-18 months will tell us which story is correct.

[Straggler Liu | AI & Semis @StragglerLiu]

Prime Agent is real and the 95.5% is real. But the framing matters.

Prime Intellect paired its open-source RLM harness with Opus 5 and ran it on ARC-AGI-3. The harness improved performance from 30% to 95.5%— slightly above the reported human expert baseline of 95.4%. The result is legitimate, but it's a harness+model score, not a pure model capability score.

ARC Prize already has Tycho at 100%, Retrodict at 99.9%. Prime Agent isn't topping the chart. What makes it interesting is the architecture.

The harness treats context as a variable and subagents as function calls inside a persistent IPython REPL. The Continual Harness stores prompts, skills, and memories as durable state the agent can read, update, and delete across sessions. That's not just a bigger context window — it's a different model of how agents interact with their own history.

The open question isn't whether 95.5% is impressive. It's whether this RLM abstraction can be productized at scale and whether it transfers beyond benchmark tasks.

[DAIR.AI @dair_ai]

Very interesting new work from Prime Intellect.

Prime Agent is an impressive open-source self-improving harnesses.

It's a harness for long-horizon agent work. A persistent IPython REPL lets the model process its own context programmatically, and a Continual Harness carries histories, memories, skills, prompts, and subagent specifications across trajectories, so improvements compound instead of resetting on every run.

Same model class, ARC-AGI-3 RHAE Best@1 moves from 30% to 95.5%. It also matches or beats native harnesses on long-context coding, GPU kernel generation, and autonomous nanoGPT speedruns.

Paper: https://t.co/HyViIlNqD5

Track more trending AI papers in our academy: https://t.co/LRnpZN7L4c

[Straggler Liu | AI & Semis @StragglerLiu]

OpenAI's Jalapeño is a statement of intent. The real question isn't TCO or throughput per MW — it's whether they can scale production and maintain competitive performance across generations. Custom silicon from hyperscalers and AI labs is the long-term margin threat to Nvidia, not just a performance benchmark.

[Straggler Liu | AI & Semis @StragglerLiu]

Credit spread widening is the canary. When credit markets start questioning off-balance-sheet AI financing commitments, equity multiples usually follow, not lead. SOX forward P/E has already dropped from 22x to 15x — the real question is whether earnings estimates have caught up to that repricing.

[Straggler Liu | AI & Semis @StragglerLiu]

You're Not Competing with Frontier Labs. You're Competing with the Companies That Own the Data.

Thomson Reuters launched Thomson 1.0 last week. The headline is a model that competes with frontier AI in legal and professional work.

The real headline: they built it for $40 million — and the final training run cost less than $450,000.

That final run was less than 1.125% of total spend. The other 98.875% went to people, data, experimentation, infrastructure, evaluation, and tooling.

Two AI Models. Two Different Questions.

Frontier labs ask: "How do we build a better general intelligence?"

Thomson asked: "How do we turn 160 years of proprietary legal content into a system that makes legal professionals faster?"

One is a moonshot — expensive, uncertain, but potentially world-changing.

The other is a model factory — replicable, cost-effective, and already generating returns.

AliPay built a business on 1.7B users. Thomson's data flywheel is the same idea—zero marginal cost to add another user with all their data.

What Thomson Actually Did

Thomson started with Qwen 3.5-397B-A17B (397B total, 17B active). They developed an intermediate checkpoint called Snowdon, then applied ~200B tokens of continual pre-training, model merging, preference optimization, and reinforcement learning. Team: ≤36 engineers. Peak compute: 368 NVIDIA B200 GPUs.

Result: 78.5 overall — trailing Opus 4.8 (79.5) but beating Gemini 3.1 Pro (78.0), GPT-5.4 (76.5), and Qwen 3.5 (73.0) on their evaluation suite.

Of 24 benchmark rows: 10 legal (41.7%), 2 tax, 1 journalism, 9 general, 2 safety. This is a Thomson Reuters-relevant composite, not a general-intelligence score.

The Real Moat Is Not the Weights — It's the System

Thomson's best results depend on proprietary legal/news retrieval, agent orchestration, citation handling, and domain workflows. Remove the tool stack and performance drops.

The moat is the combined stack: weights + content + retrieval + access + workflows + distribution.

Three Pillars of the Model Factory

The $40M model factory is a replicable paradigm for any enterprise with proprietary data and strong workflow distribution.

Data flywheel: Proprietary content generates synthetic training data → improves the model → improves workflow output → generates more data. A closed loop no generic model can replicate.

Training capability: After ~$40M invested in infrastructure and pipelines, the marginal run costs <$450K. Reusable across verticals.

Distribution advantage: The model plugs directly into existing products, reducing adoption friction and generating real-time usage data.

The Alignment Tax Is Real

Snowdon, the intermediate checkpoint, dropped from 93.55% to 78.82% on adversarial safety evaluation — a 15-point collapse. Final recovery to 90.17% remains below Qwen's 93.55%.

This is not a quality control failure. It's a structural reality: value alignment interventions can degrade safety robustness, and the loss is not fully recoverable.

The License Is a Strategy Signal

The small model is released under PolyForm Strict 1.0.0 — non-commercial only. "Open-weight" is accurate; "open source" is misleading. Strategically, Thomson gets research visibility without commoditizing the model.

What to Watch

Thomson used less than 10% of its content. Two milestones:

6 months: Did they apply the same playbook to another vertical?

18 months: Did they scale content usage to 50%+ without hitting diminishing returns?

The Bottom Line

Thomson Reuters just proved that enterprises don't need frontier labs to build world-class vertical AI. Value capture in AI is shifting from "model providers" to "data and workflow owners."

The 90% of unused content is the real upside. That's where the story goes from interesting to industry-defining.

[Straggler Liu | AI & Semis @StragglerLiu]

AI Data Center "Execution Risk" Premium: A Hypothesis Being Systematically Disproved

Morgan Stanley analyst Stephen Byrd has a straightforward thesis: AI data center construction is high-risk, so equity discount rates should be high.

It's a logical framework. The problem: the companies building these projects have spent the last four quarters systematically disproving it.

Five bitcoin mining-to-AI transitions — Applied Digital, Riot Platforms, Cipher Mining, Core Scientific, and TeraWulf — have collectively delivered something the market didn't expect: a near-perfect execution record.

Cipher Mining delivered Black Pearl two months early — at AWS's request. Core Scientific completed four of five CoreWeave campuses ahead of schedule, hitting 437MW billable capacity, up nearly 200MW in a single quarter. Applied Digital delivered 100MW at Polaris Forge on time and budget. Riot brought AMD's first 25MW online on schedule. TeraWulf's WULF Compute came in at $9.1M/MW — still within the original $8-10M guidance.

That's five companies, zero major delays, and zero cost blowouts — at a time when the broader market is convinced that execution risk is the industry's Achilles' heel.

The Scorecard

Applied Digital CEO Wes Cummins confirmed PF1 and PF2 are "progressing on time and on budget." Q3 revenue hit $126.6M (+139% YoY), adjusted EBITDA $44.1M. About 900MW under construction. PF1's 100MW delivered on time and budget in July.

Riot's initial 25MW deployment to AMD was delivered "on time and on budget" — first 5MW live in January, remaining 20MW in May. AMD exercised its option to expand to 50MW.

Cipher Mining delivered Black Pearl's initial capacity two months ahead of schedule in August — the acceleration was at AWS's request as part of its $5.5 billion, 15-year lease. Total development portfolio: 5.3GW.

Core Scientific reached 437MW billable capacity in Q2 — up nearly 200MW from Q1. Four of five CoreWeave campuses were substantially completed ahead of schedule. Q2 colocation revenue hit $137M. Build cost: $11-12M/MW.

TeraWulf's WULF Compute came in at $9.1M/MW — up from $8.6M, but still within the original $8-10M guidance range. HPC leasing revenue accounted for 71% of Q2 total revenue.

Why This Matters Now

The aggregate scorecard matters because the market's concern about execution risk has real consequences. If equity discount rates are inflated, the cost of capital for these projects is artificially high, limiting expansion and weighing on valuations. If the concern is overblown, the discount rates should narrow, unlocking significant upside.

Morgan Stanley estimates credit markets will finance more than $1 trillion in global data center spending through 2028. In 2026 alone, global AI-related debt is expected to reach $580 billion. The pricing of risk in this market — and specifically, the premium attached to execution risk — has system-wide implications. If execution risk is being systematically overestimated by even a small margin, the aggregate mispricing could be substantial.

The Industry Shift

These five companies share a common origin: they all started in bitcoin mining. In that industry, survival depends on managing extreme volatility in power costs, hardware efficiency, and asset prices. That environment breeds a survival-driven execution discipline. And that discipline appears to be translating directly into AI data center delivery capability.

Morgan Stanley noted as early as 2024 that bitcoin miners controlled ~6.3GW of large-scale operational sites with an additional 2.5GW under construction — making them "the fastest way for AI companies to access power with the lowest execution risk." Two years and four quarters of delivery data later, that thesis is being validated.

The Open Question

If these companies continue to deliver on schedule, how much should the execution risk premium narrow?

The market is pricing a risk that management teams are systematically disproving. The question for investors is not whether these companies can execute — the data already provides an answer. The question is when the valuation will catch up to the execution record.

[Straggler Liu | AI & Semis @StragglerLiu]

Goldman Sachs just dropped its Q4 CHIPS Act report. The consensus read: "China DRAM glut incoming." The data says otherwise—the gap is in the timeline, not the outcome.

Three numbers jump out:

The 7nm+ gap narrows from 92% in 2025 to 34% in 2035. Supply grows at 46% CAGR—demand at 17%. By 2035, monthly supply is 410k wafers vs 619k demand. The gap closes, but not until a decade from now.

To get there, China needs $124bn in capex from 2026 to 2035. Goldman already raised its 2030 capex forecast 79%—to $82bn. Spending is accelerating. But capital committed is not capacity delivered. Equipment still has to pass export controls. Fabs still have to run.

CXMT will cover just 50% of China's DRAM demand by 2028. Monthly capacity: 270k wafers in 2026 → 447k in 2028 → 665k in 2030. Samsung and SK Hynix together still dwarf that. In HBM, CXMT lags by 3-4 years. Not a near-term threat.

Here's what consensus misses: capacity is not supply. Yield is the hidden multiplier. Goldman assumes SMIC's advanced-node yield climbs from 23% in 2026 to 50% in 2030 and 75% in 2035. At 23% yield, half the wafers are scrap. TSMC's 7nm yield has exceeded 90% since 2018. China is not there yet, and the climb takes years.

The 2028 "peak" narrative also ignores demand. Goldman projects China's AI chip market at $678bn by 2030—69% CAGR. China DRAM demand alone hits $257bn by 2028. HBM grows at 188% CAGR.

The real question: does demand grow faster than supply can catch up? The data says yes—for at least another decade.

The market compressed a 10-year narrative into a 2-year expectation. That's the mispricing.

Source: Goldman Sachs Q4 CHIPS Act Report, Aug 24 2026.

[Straggler Liu | AI & Semis @StragglerLiu]

Alibaba's $10.3 Billion AI Bet: A Strategic Financing Gambit in the "Capital Endurance" Race

On August 23, Alibaba Group launched a massive HK$80 billion primary placement — its first new share issuance since its 2019 Hong Kong listing, and the largest primary follow-on offering in Hong Kong market history.

The details: HK$80 billion ($10.3 billion), 710 million new shares at HK$112.70 each — an 8.4% discount to the Hong Kong close and 3.6% discount to ADR levels, representing 3.7% dilution.

Alibaba will invest 100% of net proceeds into full-stack AI capabilities and infrastructure. Pro forma net cash will rise from $31 billion to over $41 billion.

The Market Reacts: Panic vs. Conviction

Hong Kong-listed Alibaba shares plunged 8.54% to HK$112.50 on August 24, briefly breaking below the placement price. The U.S. ADR fell 8.57% to $119.34.

Yet on the same day, Chairman Joe Tsai and CEO Eddie Wu bought a combined ~HK$120 million of Alibaba shares at roughly the placement price — the highest possible vote of confidence.

Why Equity, Not Debt?

Nomura pointed to high bond yields as the key reason Alibaba chose equity over debt. In a high-rate environment, issuing large-scale bonds would create rigid interest expenses, further pressuring free cash flow — which already turned negative at -$6.6 billion in the June quarter.

Equity financing, by contrast, allows Alibaba to absorb the rapid depreciation of AI hardware (3-4 year cycles) without the burden of interest payments and debt covenants. It also preserves its investment-grade credit rating.

The Hidden Logic: Arbitraging Accounting vs. Economic Depreciation

This is where the real strategic insight lies. Most investors see the placement as simply "raising cash." But there is a second-order logic at work — one that goes to the heart of how Alibaba is managing AI hardware depreciation.

AI hardware — GPUs, HBM, servers — has an accounting depreciation life of 3 to 4 years. But its economic useful life may be significantly different. If the hardware continues to generate revenue-generating capacity beyond its accounting life, the depreciation schedule overstates the true economic cost.

By choosing equity over debt, Alibaba is effectively separating the balance sheet risk from the income statement risk. Debt financing would place both depreciation and interest expense on the income statement, creating a double burden on profitability. Equity financing, by contrast, places the cost on the balance sheet as dilution — a one-time, non-cash impact that does not create ongoing interest obligations.

This structure is not just about funding AI. It is about engineering a buffer: if AI revenue growth is slower than expected, Alibaba does not face a debt spiral. If AI revenue growth is faster than expected, the dilution is quickly offset by earnings growth, and the company retains full operational flexibility.

In other words, Alibaba is betting that the economic life of AI assets will exceed their accounting life. That is the real strategic bet. The dilution is the price of making that bet without risking the income statement.

Why This Matters for ROIC — and for the Investment Case

Here is the part that most analysts have not yet priced in. The equity financing structure has a direct and measurable impact on how Alibaba's return on invested capital (ROIC) will be calculated in coming years — and therefore how the investment community will judge its success or failure.

Consider the denominator: invested capital. Under a debt-financed scenario, the denominator includes the debt principal, and interest expense reduces net income in the numerator. Under the equity-financed scenario, the denominator includes the equity capital raised, but there is no interest expense reducing net income. The dilution is not a cash cost — it is a non-cash distribution of ownership.

The accounting result is that ROIC under the equity scenario is mathematically higher than under the debt scenario, assuming identical operating performance. This is not financial engineering in the traditional sense — it is structural accounting that flows directly from the financing choice. For analysts modeling Alibaba's long-term value, the difference is material. A higher ROIC supports a higher valuation multiple, which can offset the dilution effect of the 3.7% share increase over time.

But there is a second-order effect that is equally important, but rarely discussed. The equity structure changes how Alibaba communicates with the bond market. Alibaba has a 3.6% coupon on its 2031 bonds. By not adding leverage, Alibaba sends a signal to the bond market that it intends to preserve its investment-grade rating. That keeps Alibaba's cost of debt low for any future bond issuance — a strategic option that remains open if needed.

In essence, Alibaba has chosen a financing structure that optimizes its ROIC profile, preserves its credit rating, and maintains optionality for future capital decisions. The 3.7% dilution is the entry price for all three of these benefits. Whether that is a good trade depends on whether AI revenue growth can outpace the dilution effect. The AI ARR trajectory — already exceeding RMB 49.5 billion and approaching $10 billion — is the metric to watch.

The "Capital Endurance" Phase

Alibaba's three-year RMB 380 billion AI infrastructure plan is nearly half spent. With single-quarter capex at RMB 67.7 billion (+75% YoY) and net profit down 76%, the HK$80 billion raise represents ~38% of the remaining committed spending (~RMB 190 billion).

But AI-related recurring revenue (ARR) has already exceeded RMB 49.5 billion and is expected to approach $10 billion next quarter — up from just over $5 billion in June.

A Paradigm Shift in Chinese Tech Financing

This deal signals a fundamental shift: from "relying on operating cash flow" to "raising equity capital."

Goldman Sachs estimates China's four largest tech giants will spend $102 billion on AI capex in 2026 alone. If Tencent and Baidu follow Alibaba's equity financing playbook, China's capital markets face a structural transformation in how AI is funded.

The Verdict

Alibaba's $10.3 billion placement is not just a financing event — it's a strategic signal that AI competition has entered the "capital endurance" phase.

The market is divided: short-term investors see dilution and capex pressure; long-term capital sees AI strategy, balance sheet strength, and future optionality.

BofA maintains Buy at $172. Benchmark sees $220. Michael Burry sees a stock that needs to "halve" before it's attractive again.

The answer will come in 2027 — when we see whether AI ROI can cover the cost of capital. HK$80 billion isn't the finish line. It's the fuel to keep running.

[Straggler Liu | AI & Semis @StragglerLiu]

Samsung 4nm's Turnaround: What Groq 3 Mass Production Means for the Foundry War

On August 24, NVIDIA officially announced at Hot Chips 2026 that the Groq 3 LPX inference acceleration system has entered full-scale mass production. This is not just another chip launch — it marks a potential turning point for Samsung Electronics' foundry business, which has been struggling with losses since 2022.

From 33% to 80%: Samsung's Six-Year Fight on 4nm

Samsung began mass production of its 4nm process in 2021 with yields of only about 35%. Over the next six years, the company worked steadily to improve yields and optimize the process. As recently as April 2026, Groq 3 LPU yields were reportedly still only about 33% — just one in three chips usable.

But the improvement came fast. By the time mass production launched in August, Groq 3 LPU yields had reportedly exceeded 80% — widely considered the threshold for stable advanced-node mass production. Some analysts attribute the improvement to Samsung's experience building 4nm-class HBM base dies, where it has already achieved yields above 90%.

80% is a milestone. TSMC's 4nm yields are estimated at 85-90%. The gap between Samsung and TSMC at 4nm has narrowed from a chasm to single-digit percentage points. This means Samsung can now reliably take on AI accelerator orders that demand tight reliability specifications.

Full Capacity: From Begging for Customers to Choosing Customers

Samsung's Pyeongtaek SF4 (4nm) line has been running at full capacity since late 2025. The line produces logic chips for Qualcomm and base dies for Samsung's own HBM. Groq 3 added to the tight capacity.

The direct result: pricing power has returned. In July, Samsung raised foundry prices on its SF4 process — 10% to 15% for customers in China and the US, 5% to 10% for Taiwan-based customers. This marks the first time since 2022 that Samsung's foundry business has had meaningful pricing power.

The Financial Inflection Point: Potential First Profit in 2027

Samsung's foundry business has been loss-making since 2022. But that is changing. Samsung confirmed in its Q2 earnings that foundry revenue had improved significantly, driven by HBM base dies and increased orders from US customers. The company expects double-digit year-over-year foundry revenue growth in the second half of 2026. Advanced nodes will account for more than half of foundry revenue this year, with AI and high-performance computing applications exceeding 30%, up from 15-20% at the end of 2025.

Analysts believe that if Samsung continues to raise prices, the foundry business could turn profitable as early as 2027, earlier than previously expected.

Supply Chain Beneficiaries

Groq 3's mass production benefits more than just Samsung's foundry division. Samsung Electro-Mechanics has become the largest supplier of FC-BGA (Flip-Chip Ball Grid Array) substrates for the Groq 3 LPU. FC-BGA is the high-density package substrate that connects the semiconductor chip to the motherboard. Mass production of the substrate reportedly began in Q2 2026.

At the system level, the Groq 3 LPX is a single-rack system integrating 256 Groq 3 Language Processing Units (LPUs). The LPU is the inference-specific chip technology NVIDIA acquired when it purchased Groq late last year, designed to reduce latency in AI responses by using high-speed SRAM instead of general-purpose DRAM.

Industry Implications: TSMC's Spillover Effect

According to Counterpoint data, Samsung accounted for just 7% of global foundry revenue in Q1 2026, while TSMC held over 70%.

But AI chip demand has booked most of TSMC's advanced capacity. Customers are turning to Samsung and Intel. NVIDIA chose Samsung for Groq 3. Tesla and Apple announced chip manufacturing agreements with Samsung last year. Broadcom reached an AI chip production deal with Samsung last month. Google is reportedly in talks to use Samsung's SF4 process.

The Biggest Question: The Next-Generation Groq Order

Industry observers note that whether Samsung wins the next-generation Groq LPU order is the biggest open question. If Samsung does, it would mean NVIDIA's foundry relationship with Samsung has moved from a temporary arrangement to a long-term strategic partnership — a structural challenge to TSMC's dominance.

Samsung has already set an initial wafer input target of approximately 10,000 wafers per month for Groq 3, with plans to expand to over 15,000 wafers per month by 2027 — a 50% increase.

The Bottom Line

Samsung's 4nm story is a triple reversal: yield, capacity, and pricing power. The leap from 35% to 80% yields took six years; the jump from 33% to 80% took only four months. These three signals — 80% yield, full capacity, and 10-15% price increases — point to the same conclusion: Samsung Foundry is transitioning from a technology chaser to a key node in the AI supply chain.

But the biggest variable remains unresolved. Groq 3 is Samsung's ticket into NVIDIA's AI chip foundry ecosystem. The next-generation Groq order is the renewal contract that will determine whether Samsung can hold its position in that ecosystem.

2027 is when we'll get the answer.

[Straggler Liu | AI & Semis @StragglerLiu]

The Sivers Thesis: Qualification as Moat — But the Clock Is Ticking

This is a solid piece of analysis. The core insight — that qualification may be a stronger moat than patents or raw performance — is exactly right.

Let me add two dimensions.

First, the qualification moat is real but unproven at scale.

Sivers' 8-wavelength CW-WDM MSA compliant DFB laser arrays are the foundation of its photonics business. Each channel delivers over 50mW, with 100mW designs in development. The company has demonstrated integration with GlobalFoundries' silicon photonics platform, with its laser arrays being integrated into reference designs built on GF's platform. Sivers is also a key partner in the CPO ecosystem through collaborations with Ayar Labs and others. Win Semi provides high-volume manufacturing capacity without requiring billions in capex.

But here is the gap: qualification in a reference design is not the same as being locked into a high-volume production supply chain. The switching cost thesis is valid — once a laser array is integrated into a qualified optical engine design, swapping suppliers becomes complex and expensive. But that lock-in only matters if volume materializes. Sivers' opportunity pipeline grew 77% to $799 million in Q1 2026, but net sales fell 22% year-over-year to 61.9 million Swedish kronor, impacted by US government shutdown delays and FX headwinds. The pipeline is a leading indicator. Revenue is the lagging one that actually pays the bills.

Second, the M&A speculation is logical but premature.

The comparison to Lumentum, which scaled via acquiring Oclaro, NeoPhotonics, and others, is apt. As Sivers' revenue grows and its market cap expands, M&A could become a realistic path to vertical integration. The company is already evaluating a potential dual listing on Nasdaq New York, which would enhance access to US capital markets and potentially enable future strategic acquisitions. But for now, it is speculation. The company first needs to convert that $799 million pipeline into real revenue. 2027 remains the critical year.

The bottom line: the technology is real. The moat logic is sound. The partnerships are in place. But Sivers is still in the show-me phase. Qualification as a reference design partner is a necessary first step, but it is not the same as being the sole source in a shipping product. The market is pricing the potential. The earnings need to follow.

[peakybulls @peakybulls]

$SIVE Teil 2/2
Ich bin bei $SIVE noch eine Stufe tiefer in die DD gegangen.

Mein wichtigstes Learning:
> Der potenzielle Moat ist nicht einfach: "Sivers macht gute Laser"
> Der eigentliche Moat könnte tiefer liegen.

> Sivers liefert keine Standardlaser.
> Die DFB Laserarrays sind auf die Integration mit Silicon Photonics optimiert.

> 8-Wellenlängen Arrays.
> Hohe Leistung (Laser liefert genug Lichtenergie).
> CW WDM (Wavelength Division Multiplexing)
> Passive Alignment.
> Flip-Chip Integration mit Silicon Photonics
> High-Yield Manufacturing.

> Bei AI optics kauft ein Kunde nicht einen einzelnen Laser.
> Der Laser muss in ein komplettes System passen.
> Sivers muss also einen Laser liefern, der so entwickelt ist, dass er ins entsprechende System integriert und qualifiziert werden kann.
- Wellenlänge.
- Power.
- Thermik.
- Packaging.
- Alignment.
- Coupling.
- Reliability.
- Yield.

> Alles muss zusammen funktionieren.

> Der potenzielle Switching Moat entsteht damit möglicherweise durch die Qualification.
> Nicht nur durch Patente. Nicht nur durch Performance.

> Wenn ein Laserarray bereits in einem Optical Engine, einem Silicon Photonics Design oder einer Reference Platform integriert und qualifiziert ist, wird der Wechsel des Suppliers deutlich komplexer.

> Dann kommt der nächste Punkt: Scale.

> $SIVE hat mit Win Semi einen High-Volume Manufacturing Partner für die DFB laserarrays an der Seite.

> Win Semi gibt Sivers heute Flexibilität. Scale ohne Milliarden-Capex.
> Sivers kann Technologie und Nachfrage erst wachsen lassen.
> Und später entscheiden, wie viel der Wertschöpfungskette man selbst besitzen möchte.

Hier wird M&A interessant:
> $LITE ist durch M&A zu einem deutlich breiteren Photonics Unternehmen geworden:
- Oclaro
- IPG Telecom Transmission
- Cloud Light
- NeoPhotonics

> $SIVE muss deshalb nicht für immer nur ein spezialisierter Laserhersteller bleiben.

> Mit wachsender Revenue. Höherer Market Cap. Nasdaq Listing...
> Könnte Sivers später selbst M&A durchführen.
- Laser-IP.
- Packaging.
- Optical Engines.
- Manufacturing.
- Test.

> ABER: Noch ist das reine Spekulation.
> Revenue muss zuerst folgen.
> 2027+ bleibt entscheidend.

> Ich bin gespannt auf die Earnings!

[Straggler Liu | AI & Semis @StragglerLiu]

2026 World Robot Conference: Humanoids Just Broke Records — Now the Real Work Begins

A humanoid robot just ran 100 meters in 9.32 seconds. The same robot jumped 2.8843 meters — higher than any human has ever jumped. Another won the 400 meters in 38.15 seconds, beating the human world record by nearly five seconds.

At the 2026 World Robot Conference in Beijing, 373 exhibitors from 26 countries showcased over 3,000 products, with 311 new products making their debut.The week also featured the second World Humanoid Robot Games, with 666 teams and 2,056 robots competing in 51 events across 1,301 matches.Unitree Robotics went public during the conference.The message couldn't be clearer: humanoid robots have arrived — but not in the way most people think.

The Real Signal Isn't the Records — It's the Commercial Push

The headline numbers are staggering. China shipped over 40,000 humanoid robots in H1 2026, accounting for 97% of global shipments, up from 84.7% a year earlier. The industry is growing at over 20% annually, with 2025 revenue exceeding RMB 300 billion, according to the Ministry of Industry and Information Technology.

But the deeper signal is what those robots are actually doing.

The conference marked a decisive shift from "demonstration" to "deployment." Ubtech demonstrated nearly 10 humanoid robots working in coordinated factory workflows — handling automotive sheet metal parts, machine parts, and box stacking. Galaxy General Intelligence brought https://t.co/7PviBN4N20's front warehouse into the exhibition, with robotic arms autonomously navigating shelves for picking and packing. Galbot's S1 robots are already deployed at CATL, Bosch, and major automakers — doing real work, not just dancing.

Xu Xiaolan, chairwoman of the Chinese Institute of Electronics, said embodied intelligence products are at a critical inflection point — moving from "small-batch trials" to "large-scale deployment." As one exhibitor put it: "Last year, more robots were dancing. This year, more are working."

Why China? Scale, Cost, and Supply Chain Density

The 97% figure is not a coincidence. It's the result of three converging forces:

First, the supply chain. Every component — motors, reducers, sensors — can be sourced within a three-hour drive of the Yangtze or Pearl River deltas. A joint motor that cost 50,000–60,000 yuan in 2018 now costs 500–600 yuan. The localization rate of core components has climbed from under 50% to over 90%.

Second, price. Unitree's G1 retails at $5,600. Tesla's Optimus is still targeting $25,000–$30,000. At an order of magnitude price difference, the market naturally tilts toward the lower-cost option.

Third, policy. "Embodied intelligence" was written into China's government work report in 2025 and is a priority in the 15th Five-Year Plan. Over 70 embodied-intelligence training grounds are already operating, according to the China Academy of Information and Communications Technology.

The Competitive Landscape

Tesla's Optimus V3 is expected to enter mass production in H2 2026, sourcing about 70% of its core components from over ten Chinese suppliers. Boston Dynamics' Atlas and Figure AI's humanoids are still in the hundreds-of-thousands-of-dollars range.

In H1 2026, Chinese manufacturers took over 97% of global shipments. Zhiyuan Robotics shipped approximately 9,700 units (43% global share), Unitree shipped over 7,000 units (31%). The two companies alone accounted for about 75% of all humanoid robots shipped worldwide. Tesla, Figure AI, and Agility Robotics were left far behind.

But shipment volume alone doesn't win the long game. The real competition is unfolding on two other fronts.

What Actually Matters for the Next 3-5 Years

First, the data bottleneck. Training a general-purpose robot brain requires millions of hours of real-world data. The industry currently has roughly one-tenth of what it needs — a hundredfold gap. Unlike LLMs, which can scrape the entire internet, robots need physical interaction data — and that data is expensive, slow, and hard to collect at scale. OpenAI CEO Sam Altman recently acknowledged: "we have no idea what to do."

Second, the "ChatGPT moment" timeline. The industry is deeply divided on when robots will achieve general-purpose capability. Unitree CEO Wang Xingxing projected a 2028 "ChatGPT moment" when robots can handle 70–80% of everyday tasks without specialized training. ACE Robotics chairman Wang Xiaogang is more optimistic — end of 2027. NVIDIA CEO Jensen Huang believes "the ChatGPT moment for robots occurred years ago." The wide divergence tells you how uncertain this still is.

Third, the real deployment gap. Despite all the headlines, humanoid robots are still overwhelmingly used for entertainment and research — not real work. Counterpoint data shows that in H1 2026, 33.6% of shipments went to entertainment and performance, and 27% to data production and research. Together, that's over 60% of all robots shipped. Service and guidance accounted for about 19%, and industrial/manufacturing less than 20%. The industry is still selling into categories where failure is acceptable and economic returns are secondary.

The real test will be when that ratio flips — when industrial and service applications exceed 50% of shipments. That's when the economic case becomes self-sustaining.

The Technology That Actually Matters

The 100-meter and high-jump records are spectacular, but they're not the real story. Balance and speed are table stakes now. The real frontier is generalization — can a robot trained in one factory work in another? Can it adapt when you move a chair? Can it handle novel objects without crashing?

This is where the industry is investing heavily. Foundation models for robotics are evolving rapidly. Yaskawa has installed adaptive AI on its robots that eliminates manual adjustment for different tasks. Galaxy General Intelligence's "Galaxy Star Brain" allows robotic arms and humanoid robots to share the same "brain." The embodied intelligent robotic arm is, in effect, "an old employee with a smart brain."

The Beijing Humanoid Robot Innovation Center released the "Huisikaiwu" embodied intelligence platform, integrating the State Grid's Guangming large model to form a "large brain + small brain" collaborative architecture, already deployed at multiple substations. Ubtech has implemented its "Swarm Brain Network 2.0 + Co-Agent" AI dual-cycle system, enabling collaborative training of multiple humanoid robots in real factory environments.

Better simulation, world models, and real-world data collection are starting to close the generalization gap.

Where the Industry Is Headed

Deutsche Bank has dramatically revised its humanoid robot shipment forecast upward — from 17,500 units in 2025 to nearly 50,000 in 2026. By 2030, they project ~700,000 units. By 2050: 7 million. Goldman Sachs sees the next phase of AI shifting from chips to real-world deployment, with humanoid robots as the "clearest monetization frontier."

But the economics are still unproven at scale. Unitree's $5,600 G1 is a breakthrough price point, but can it do useful work? The answer depends on whether the software can catch up to the hardware.

The Bottom Line

The 2026 World Robot Conference wasn't about robots breaking human records. It was about an industry crossing a threshold — from "can we build it?" to "can we deploy it at scale and make money?"

China has won the first round: manufacturing, cost, and supply chain. The US still leads in foundational AI research and software. Europe is strong in industrial robotics and precision engineering. The next phase will be won by whoever solves the generalization problem first — whoever can make a robot that works reliably in any environment, not just the one it was trained in.

The robots are here. They can run faster and jump higher than any human. Now they need to learn how to do useful work. That's the real race — and it's just getting started.

[Straggler Liu | AI & Semis @StragglerLiu]

CRDO Earnings Preview: A $43B, 91x P/E Connectivity Company — Can It Deliver the Optical Inflection?

On September 1, 2026, after market close, Credo Technology Group (NASDAQ: CRDO) will report Q1 FY2027 earnings. This $43 billion market cap, ~91x P/E company is at a critical validation point.

The Growth Trajectory in One Chart

Over the past 12 months, Credo has executed a remarkable financial transformation: FY2024 revenue of $193M→ FY2025 $437M→ FY2026 over $1.3B, up 206% year-over-year. Non-GAAP net income surged from ~$100M in FY2025 to $662M in FY2026, up over 5x. Q4 FY2026 alone delivered $437M in revenue, up 157% year-over-year, with non-GAAP net income of $226.7M — a 51.9% net margin.

The driver is explosive AI infrastructure demand — Credo's Active Electrical Cables (AECs), retimers, and optical connectivity products are critical components in AI data center scale-out networks.

Q1 FY2027 Guidance: Still Impressive, But Slowing Sequentially

Credo's Q1 FY2027 guidance, issued on June 1, 2026:

MetricGuidance RangeRevenue$465M – $475MGAAP Gross Margin66.9% – 68.9%Non-GAAP Gross Margin67% – 69%Non-GAAP Opex$86M – $90MDiluted Weighted Avg Shares~199M
Q1 revenue implies 108%-113% year-over-year growth (vs $223.1M in the same quarter last year). But sequentially, Q4 FY2026 revenue was $437M, and Q1 guidance midpoint is $470M — only ~7.5% sequential growth. This is a stark contrast to the >50% sequential growth rates of previous quarters.
Management explicitly stated on the earnings call: FY2027 first half will see only "mid-single-digit" sequential growth, with the inflection coming in the second half.
The Inflection Engine: Optical Business
Credo's growth story is transitioning from "copper connectivity" to "optical connectivity."
FY2027 full-year revenue growth guidance is >80%. The core driver is **optical products expected to contribute over $600M in revenue**, with ZeroFlap optical transceivers, silicon photonic PICs, and optical DSPs each expected to contribute over $100M. Management expects roughly half of FY2027 absolute dollar revenue growth to come from optical products and half from copper products.
On May 28, 2026, Credo completed its $750M acquisition of DustPhotonics. DustPhotonics brings industry-leading silicon photonic PIC technology covering 800G, 1.6T, and 3.2T near-packaged optics (NPO) and co-packaged optics (CPO). Credo has now built a vertically integrated connectivity stack: SerDes → DSP → silicon photonics → system integration.

Three Critical Questions for September 1

Question 1: Is the optical inflection on track?

This is the biggest question for the September 1 earnings report. Management guided H1 to only mid-single-digit growth, with a significant acceleration in H2. The Q1 report will be the first look at optical product revenue contribution post-DustPhotonics integration. Investors need to watch: ZeroFlap optical transceiver customer adoption progress, silicon photonic PIC production timelines, and optical DSP design-win conversion. If optical revenue falls short, the FY2027 >80% growth target could be at risk.

Question 2: Is customer concentration risk improving?

This is Credo's most discussed structural risk. Historically, the top three customers accounted for 39%, 32%, and 17% of revenue respectively — 88% total. But the company is improving this structure: four hyperscale customers now exceed the 10% revenue threshold, and a fifth has secured a design win. The September 1 report will reveal whether customer concentration is continuing to improve.

Question 3: Is the valuation already pricing in the optimism?

Current stock price is ~$222.61, market cap ~$43B, P/E ~91x. Wall Street's 19 analysts polled by S&P Global have a consensus "Strong Buy" rating with an average price target of $283.23. But there is significant divergence: Stifel has a $350 target, while Rosenblatt maintains a "Neutral" rating with a $215 target. Bulls see the optical inflection driving >80% FY2027 growth; bears worry about customer concentration, high valuation, and the timing risk of optical revenue realization.

Conclusion

Credo is at a critical narrative inflection point. The past 12 months of financial performance have proven the success of its AEC and retimer business. But the current market pricing (91x P/E) reflects expectations for a successful optical business transition — a transition that has not yet been fully validated in the financial data.

The September 1 earnings report will be the first look at optical product contribution post-DustPhotonics integration, and will validate whether the H2 inflection remains on track. If optical revenue beats expectations, Credo could break out of its current valuation range. If it falls short, the 91x P/E multiple will face pressure.

Credo's core investment thesis is: can a company transitioning from "copper connectivity" to "optical connectivity" evolve from an AEC leader to a full-stack connectivity platform in the high-growth AI data center interconnect market? September 1 will provide the first critical data point.

[Straggler Liu | AI & Semis @StragglerLiu]

Nvidia's options market just signaled something the stock market hasn't fully priced in.

According to Reuters, options traders are pricing in a $280 billion swing in Nvidia's market value after Wednesday's Q2 earnings — a 5.4% move in either direction.

That's below the 6.5% implied move ahead of its May report, and significantly below the 7.4% historical average over the last 12 quarters, according to ORATS.

The signal here isn't the magnitude — it's the direction of the trend.

As Matt Amberson, founder of ORATS, put it: "That shows some complacency for Nvidia, and it means it's getting more predictable".

Chris Murphy, co-head of derivatives strategy at Susquehanna, went further: "I think the beginning of the AI era when Nvidia was surprising everybody with the huge earnings beats and 10, 15, 20 percent moves, that's kind of over".

What does this mean?

First, normalization is underway. The options market is pricing Nvidia as a mature growth stock, not a meme-like volatility machine. The 5.4% implied move — while still representing $280 billion in market cap, more than 90% of S&P 500 constituents — reflects a market that expects predictability, not fireworks.

Second, the bar for surprise has been raised. To move Nvidia 10%+ post-earnings, the company would need to deliver something truly extraordinary. The market is no longer assuming that.

Third, this is happening against macro headwinds. Nvidia has fallen for seven consecutive trading days, 30-year Treasury yields hit a 19-year high last week, and the broader tech sector is under pressure. The options market is pricing in a contained outcome despite this uncertainty.

The takeaway for investors:

The AI trade is maturing. Nvidia's earnings are no longer a binary "blowout or bust" event. The market is pricing in predictability — which means the risk-reward profile is shifting.

If Nvidia delivers a solid beat, the upside may be capped by elevated expectations. If it disappoints, the downside could be larger than the options market implies — because the "complacency" Amberson identified cuts both ways.

The options market is telling us the AI era's 15% earnings swings are over. The question is: has the stock market fully absorbed that yet?

[Reuters @Reuters]

Nvidia shares set for $280 billion price swing after earnings, options show https://t.co/Ldr6vPgh4O https://t.co/Ldr6vPgh4O

[Straggler Liu | AI & Semis @StragglerLiu]

Broadcom's $100 Billion AI Bet Is Creating "Phantom Leverage" — and the Credit Market Is Starting to Notice

Broadcom has a problem that most chipmakers don't: its customers can't afford its chips.

Anthropic wants to buy Broadcom's custom AI accelerators. But the cost of building out AI infrastructure at scale — data centers, networking, cooling, power — has become so enormous that even well-funded AI labs can't write the check upfront. So Broadcom did what any rational chipmaker would do: it started financing its own sales.

The deal is simple. Broadcom sets up a Special Purpose Vehicle, which borrows money from Apollo and Blackstone. The SPV buys Broadcom chips and leases them to Anthropic. And Broadcom guarantees the debt — roughly 85% of the senior tranche. In June 2026, the first $35 billion deal closed. Now Broadcom is in talks for another $60 billion to $70 billion, potentially pushing the total toward $100 billion.

The equity market loves this. AI semiconductor revenue reached $10.8 billion in Broadcom's latest quarter, with management seeing line of sight to more than $100 billion in AI chip revenue alone by 2027. The stock trades at 65 times trailing earnings. The growth story is intact.

The credit market sees something else.

The Signal from the Bond Market

Broadcom's five-year Credit Default Swaps — insurance against default — jumped 28 basis points in August alone. That's a larger increase than both Oracle and SpaceX. Its 5.15% bonds due 2031 saw yields rise 14 basis points over the same period.

A 28-basis-point move in CDS in a single month is not noise. It's a signal that the bond market is repricing Broadcom's credit risk — not because of its operating performance, but because of the guarantees it's writing. As Tony Trzcinka, investment grade portfolio manager at Impax Asset Management, told Bloomberg: "The rise in Broadcom's CDS seems to me is more specific to their balance sheet than overall angst over AI investment".

The concern is not that Broadcom is weak. The concern is that Broadcom is concentrating risk.

The Structural Fragility

JPMorgan strategist Tarek Hamid described this as "phantom leverage" — a growing layer of obligations that don't show up as normal debt on a company's balance sheet today, but could become very real in a downturn. Leases, purchase commitments, residual value guarantees — all are accumulating underneath the AI ecosystem. Hamid warned that these backstops could stretch into the trillions.

The key insight is this: the guarantees are most dangerous precisely when they would be needed most.

If AI monetization disappoints and Anthropic slows investment, the SPV becomes fragile. Broadcom would have to honor billions in guarantees — at exactly the same time it would be facing weaker chip sales and heavier supply commitments. The guarantee becomes a call on Broadcom's balance sheet at the worst possible moment.

The ASIC Asymmetry

There's a second layer to the risk that makes Broadcom's situation different from Nvidia's.

Broadcom's AI accelerators are custom ASICs, designed for specific customers. Google's TPUs are built for Google's workloads. Anthropic's chips are built for Anthropic's models. Unlike Nvidia's relatively standardized GPUs, these chips have little to no secondary market. If a customer disappears or suddenly cuts demand, the residual value of those chips collapses.

This makes the residual-value assumptions behind the financing structures even more critical. The SPV's debt is priced on the assumption that the underlying chips have value beyond the initial lease. If that assumption fails, the entire financing structure frays.

The Capital Markets Map

Who gets paid: Broadcom gets AI chip revenue it otherwise couldn't book. Apollo and Blackstone get investment-grade yields on private credit. Anthropic gets compute capacity without upfront capex.

Who bears the risk: Broadcom's balance sheet. The guarantees don't show up as debt, but they are contingent liabilities. In a downturn, they become real.

The valuation gap: The equity market is pricing Broadcom's AI growth story at 65x trailing earnings. The credit market is pricing the contingent liabilities through widening CDS. One of these markets is wrong.

The invalidation trigger: If Anthropic's revenue growth slows, if AI capex from hyperscalers cools, or if a major customer diversifies its chip sourcing (as Google just did with Marvell), the guarantees start to look more like liabilities than assets.

The Takeaway

Broadcom has built a brilliant financial machine. It sells chips, finances the purchase, guarantees the debt, and books the revenue. The AI buildout gets funded. Everyone wins — until someone doesn't.

The credit market is starting to price the possibility that someone doesn't. The equity market hasn't caught up yet.

The question for Broadcom shareholders isn't whether AI demand is real. It's whether a company that trades at 65 times earnings can absorb billions in contingent liabilities if that demand falters.

The bond market is starting to wonder. The stock market isn't — yet.

https://t.co/JtQeVxXyvi

[Straggler Liu | AI & Semis @StragglerLiu]

Hot Chips 2026: Micron Quantified the Memory Wall — 3x Compute, 2x Memory, 90% Area, and 17% of Llama 3 Training Failures

On August 23 at Stanford University's Hot Chips 2026, Micron HBM design architecture fellow Raghu Sreeramaneni did something unusual. Instead of selling HBM4's speed, he quantified exactly how far AI has pushed memory — and how far memory is falling behind.

The compute gap: 3x.

Micron stated that AI accelerator compute performance roughly triples every two years. Over the same period, HBM bandwidth grows by less than twofold. The gap is not closing — it's widening. The memory wall is not a metaphor. It's a measurable, accelerating divergence.

The footprint: 90% memory.

In a typical system-in-package integrating two GPUs with eight 12-layer HBM4 stacks, memory accounts for roughly 90% of the semiconductor area. Memory occupies more than eight times the area of the compute silicon. The "compute engine" is a minority in its own package. As Sreeramaneni put it: "HBM is smaller than a postage stamp, yet it requires design, process, and packaging technologies to all perfectly match — it's the most difficult product."

The wafer cost: 3x DDR5.

Producing the same memory capacity with HBM requires roughly three times as many wafers as standard DDR5 DRAM. With every generation — faster speeds, larger dies, more stacking layers — that gap is growing. HBM delivers bandwidth, but at a steep silicon efficiency cost. For the memory industry, this means HBM output per fab is significantly compressed.

The reliability cost: 17% of Llama 3 interruptions.

Micron cited Meta's Llama 3 training data, noting that HBM accounted for roughly 17% of unintended training interruptions. Over a 54-day training run across thousands of GPUs, memory became a leading cause of system failure. This is the hidden cost of pushing memory to its limits: not just performance, but reliability under sustained, extreme workloads.

The thermal paradox.

Sreeramaneni also noted that as data transfer speeds increase, thermal management has become a primary design consideration. HBM is a multi-layer DRAM stack. The bottom layers generate the most heat. Cooling happens from the top. Heat must travel through the entire stack to escape. Thinner chips dissipate heat worse, as the proportion of oxide rises and oxide conducts heat far less effectively than silicon. The industry is now designing chips around thermal constraints first.

The takeaway.

Micron's presentation wasn't about HBM4's specs. It was about the fundamental physics of memory scaling in the AI era. The numbers tell a consistent story: compute is accelerating faster than memory can keep up; memory consumes more area than the compute it serves; memory costs more silicon than any other component; and memory is now a leading cause of system failure at scale.

Micron's HBM4 has begun shipping to Nvidia for the Vera Rubin AI chip. But Sreeramaneni's warning is structural: this is not a bandwidth problem — it's a physics problem. Area, cost, heat, reliability, and bandwidth are locked in a self-reinforcing loop that no single generation of HBM can break.

The memory wall isn't a metaphor. It's a set of numbers. And those numbers are getting worse.
https://t.co/vJavgS0qst

[Straggler Liu | AI & Semis @StragglerLiu]

The 775-Micron Ceiling: SK Hynix's Physical Limit Acknowledged at Hot Chips 2026 Is Redrawing HBM's Competitive Boundaries

At Hot Chips 2026, SK Hynix packaging engineering VP Jaesik Lee delivered a presentation on advanced packaging for HBM. Like most Hot Chips talks, it was packed with technical details: TSVs, micro-bumps, wafer thinning, stacking and underfill. But hidden behind those slides was a more fundamental story — physics and commercial reality are colliding head-on, and SK Hynix just made a set of strategic decisions that will shape AI infrastructure for the next three to four years.

One Number, One Limit
Lee kept returning to one number throughout his presentation: 775 microns.

This is the Z-height ceiling for HBM4 packaging. The JEDEC HBM4 standard raised the package thickness ceiling from 720 microns, which held through HBM3E, to 775 microns, easing the pressures associated with adopting hybrid bonding. But this number wasn't chosen arbitrarily — 775 microns is exactly the standard thickness of a 300mm logic wafer.

Why does this alignment matter? Because HBM and the GPU sit side by side in the same package, sharing the same mechanical and thermal constraints. The HBM stack cannot exceed the height of the logic chip beside it — otherwise the entire package's mechanical integrity and thermal path break down. This is a hard constraint, not a soft one that can be optimized away.

Beyond 16 Layers, Every Additional Layer Gets Harder
Inside this 775-micron box, SK Hynix is doing one thing: packing more DRAM layers into less space.

SK Hynix has 12-Hi HBM4 in mass production and has moved 16-Hi HBM4 into customer qualification. The 16-layer stack holds 48GB, but the cost is that each core die must be thinned to about 50 microns — thinner than a sheet of paper (about 100 microns).

Compared to 12-Hi designs, the 16-Hi design reduces die thickness to 0.9x, gap height to 0.5x, and bump pitch to 0.9x. Data from SK Hynix's presentation shows HBM's thermal burden has risen 2.2x compared to early generations. The drivers are clear: pin speeds climbed from 1 Gbps in early HBM to 8 Gbps in HBM4, concentrating more power in the same area, while stack counts double every two generations. The combination of bandwidth growth and layer count increases is turning thermal management into an increasingly severe challenge.

There's a more subtle problem: thinner dies actually dissipate heat worse. When chips are ground to 50 microns, the proportion of oxide in the stack rises, and oxide conducts heat far less effectively than silicon. In other words, you're working to squeeze more heat sources into a smaller space while the pathways for heat to escape become less efficient. This is an inflection point that is rapidly approaching.

Hybrid Bonding Pushed to HBM5 — A Strategic Decision
Lee made an announcement during his presentation that forced the entire industry to reassess its roadmaps: hybrid bonding will not be used for HBM4E and will arrive at HBM5 at the earliest.

This is not a technology failure — it is a strategic judgment about physical limits. Counterpoint Research expects hybrid bonding to enter mass production around 2029 to 2030, aligning with the HBM5 launch timeline.

This means MR-MUF must survive through both HBM4 and HBM4E.

MR-MUF (Mass Reflow-Molded Underfill) and TC-NCF (Thermo-Compression with Non-Conductive Film) are the two main bonding approaches for HBM stacking. TC-NCF offers high productivity and low thermal resistivity but is sensitive to chip warpage, while MR-MUF handles thin-die warpage better at the cost of higher thermal resistivity and a narrower gap-fill window. SK Hynix has chosen MR-MUF as its primary approach and pushed it to 16-Hi HBM3E with new warpage control and fine-pitch gap-fill techniques.

But 16 layers are already pushing the limits of MR-MUF's capabilities. The 16-Hi design shrinks gaps by half and thins dies to under 50 microns. Manufacturing teams must fill gaps that have shrunk by half while controlling warpage in dies thinner than 50 microns. Warpage can misalign the stack, damage connections, and reduce production yield. The 775-micron limit gives MR-MUF more time, but it also means SK Hynix is betting that one technology can survive the next three to four years of product cycles.

Hybrid bonding theoretically solves these problems. Compared to MR-MUF, hybrid bonding can increase core die thickness by 24% — because it eliminates the space taken by micro-bumps — reduce TSV pitch to under 18 microns, and cut thermal resistance by roughly 35% while enabling 20+ layer stacks.

But the gap between "theoretical" and "production-ready" is vast. Lee's decision essentially said: we'd rather squeeze one more generation out of MR-MUF than take a risk on immature hybrid bonding.

Samsung Chose a Different Path
On the same Hot Chips stage, Samsung DRAM design team's Sangwook Han presented a completely different roadmap.

Samsung's roadmap unfolds in three phases, culminating in an architecture called zHBM — DRAM stacked directly on top of compute chips like GPUs and TPUs, completely eliminating the 2.5D interposer. This isn't an evolution of 2.5D packaging — it's true 3D integration.

The numbers Samsung presented are striking: versus a standard HBM4E stack, zHBM claims 70% higher power efficiency and 230% greater DRAM bandwidth, saving 100W per DRAM module while freeing up 8.3% of additional power headroom for the GPU.

Before zHBM, Samsung is also advancing cHBM (custom HBM), adding SoC-like functionality to the base die through advanced logic processes, potentially offloading workloads from the XPU and freeing 5-10% of GPU area for more compute.

SK Hynix chose to squeeze more out of an existing architecture. Samsung chose to reinvent the architecture. Two roads, two bets. Which one is right? Given the reality that hybrid bonding has been pushed to 2029-2030, Samsung's zHBM roadmap looks like a more aggressive long-term bet, while SK Hynix is using MR-MUF to hold its ground in the near-to-medium term.

Will the 775-Micron Ceiling Be Broken?
Not in the short term. But longer term, the ceiling may be pushed higher.

Some reports suggest JEDEC is discussing relaxing the HBM height standard from 775 microns to approximately 900 microns. If this discussion bears fruit, it would create more room for hybrid bonding and higher stacks. But even if the height standard is relaxed, a more fundamental problem remains: every DRAM layer must be ground thinner, and thinner chips dissipate heat worse. This is a materials science problem, not a standards problem.

Samsung's zHBM offers an alternative solution — bypassing the physical limit through architectural change rather than continuing to compress within the existing architecture. Eliminating the 2.5D interposer removes a major source of thickness and thermal resistance. But 3D stacking brings its own challenges: heat conducts directly from DRAM to the compute chip below, potentially causing the compute chip to overheat.

Capital Markets Impact
This technology decision is not an academic discussion. It has clear, quantifiable capital markets implications.

Short term (2026-2028): MR-MUF supply chain benefits. SK Hynix has begun supplying 12-layer HBM4E samples using advanced MR-MUF, with thermal resistance improved by approximately 17% over HBM4. SK Hynix holds approximately 70% of Nvidia's Vera Rubin HBM orders — meaning MR-MUF packaging equipment and materials suppliers will continue to benefit for the next three to four years. Samsung is also advancing cHBM4, but its Heat Path Block technology covers 50% of PHY area and can reduce peak temperature by over 35% — a different technological path, but both paths require significant packaging capacity in the near term.

Medium term (2028-2030): The 16-layer ceiling becomes a performance bottleneck. The hybrid bonding delay means 16-layer stacking will be the effective ceiling for the next three to four years. The memory bandwidth and capacity growth rate for AI chips may fall short of expectations — this is not a demand problem, but a physical limit. If Samsung's zHBM progresses as planned, it could gain a significant technology lead in the 2028-2030 timeframe.

Long term (2030+): The company that breaks through first will capture enormous premiums. The manufacturer that first achieves mass production of hybrid bonding or 3D integration will reap both technology premiums and market share gains. SK Hynix chose to continue optimizing MR-MUF while positioning hybrid bonding for HBM5; Samsung chose to bypass 2.5D packaging limitations through zHBM. Which path succeeds first will determine the HBM market landscape for the next five years.

Conclusion
775 microns is a physical fact, not a problem that can be "optimized away." It determines how high HBM can stack, how fast it can run, and how much heat it can dissipate. The decision SK Hynix made at Hot Chips 2026 — pushing hybrid bonding to HBM5 — is an acknowledgment of this physical fact, and a systematic judgment about its own technology roadmap and competitive positioning.

The consequence of this decision: MR-MUF must survive the next three to four years of product cycles, 16-layer stacking may become the effective ceiling during this period, and Samsung's zHBM roadmap offers a completely different path. HBM competition is no longer just about bandwidth and capacity — base die architecture and advanced packaging are becoming the decisive battleground.

Physics doesn't care about roadmaps. It doesn't care about market caps either. But investors should.

https://t.co/sBTSSHeEXG