Etched

Phase A — Understand the business

Company Overview

What it is. Etched builds Sohu — an inference-only ASIC that hardwires the transformer architecture (specifically the attention mechanism) directly into silicon as fixed-function logic rather than running it as software on programmable cores. The entire company is one bet stated in one sentence by CEO Gavin Uberti: "almost every major breakthrough was simply a transformer applied to a new domain" → so build silicon for nothing but transformers, trade all flexibility for a step-change in throughput and cost.

Business model. Not a merchant-chip vendor selling loose parts. Etched sells full systems — branded "Frontier Inference Clusters" (FIC): custom chips + racks + cold plates + networking + a co-designed software stack, sold as an integrated appliance. The economic pitch is tokens-per-dollar and tokens-per-watt at the rack level, not $/chip. Uberti's framing of the prize: "whoever produces the most tokens will be the most valuable company in the world".

Products. One product line: Sohu (the chip) and the FIC rack it ships inside. A Sohu Developer Cloud (interactive playground / preview) is promised alongside first hardware. No second silicon generation is publicly named — references to "Sohu-2" in the wild appear to be conflation, not a disclosed roadmap part.

Customers / suppliers / competitors (detail in Lens 2/3/7). Customers: undisclosed, but $1B+ in signed forward contracts claimed pre-shipment. Foundry: TSMC (N4P). Memory: HBM3E (SK Hynix among suppliers per hiring provenance). IP/interconnect partner: Rambus (memory-interface IP). Competitors: NVIDIA (the incumbent it is aimed at), plus the inference-ASIC cohort — Groq, Cerebras, SambaNova, Taalas, and the hyperscaler silicon (AWS Trainium/Inferentia, Google TPU, Microsoft Maia, Meta MTIA).

Contract structure. The $1B is described as forward contracts / customer commitments for Sohu-powered systems, i.e. pre-revenue bookings, not recognised revenue. Terms, deposits, take-or-pay conditions, and cancellation rights are not disclosed — n/a — private, not disclosed. This matters enormously (see Lens 13): a "booking" from a friendly early customer with a walk-away clause is worth a fraction of a take-or-pay PO.

---

Supply Chain

Map (upstream → Etched → end customer), every named stakeholder:

• EDA / IP. Rambus supplies memory-interface (HBMHigh-bandwidth memoryMemory stacked in layers beside a processor so data reaches it faster. The usual bottleneck in AI chips. controller/PHY) IP — publicly confirmed partnership. Standard cell / physical-design tooling (Synopsys/Cadence) implied but unnamed.
• Foundry — the single most important node. TSMC N4P (4nm), built at reticle-limit die size (~800mm²) — the largest single-exposure die possible. A0 silicon returned with claimed first-pass success. Chokepoint: TSMC leading-edge capacity is the gate on every ambition here; a reticle-limit die on N4P competes for the same wafers and CoWoS-class advanced packaging as NVIDIA, AMD, Broadcom-designed ASICs, and every hyperscaler part. Etched is a tiny buyer against those whales.
• Strategic alignment with the foundry. VentureTech Alliance (a fund affiliated with TSMC) invested in the Series B. This is a genuine signal — foundry-linked capital de-risks (a little) the capacity-access problem that kills most fabless startups. It does not guarantee priority allocation.
• Memory. HBM3E, 144GB per chip. HBM3E is itself a supply bottleneck controlled by SK Hynix / Samsung / Micron; Etched has hired from SK Hynix. Single-source risk: HBM allocation is oversubscribed by NVIDIA/AMD — a small startup's HBM supply is a real constraint, not a footnote.
• Advanced packaging. A reticle-limit compute die + 144GB HBM3E requires CoWoS-class 2.5D packaging (also TSMC-gated, also oversubscribed).
• Assembly / bring-up. Etched opened a Taiwan manufacturing/engineering facility for 24/7 bring-up and runs a 2MW datacenter in its San Jose office for internal validation.
• Etched (fabless design house). ~400+ engineers, hired from NVIDIA, Google TPU, Broadcom, SK Hynix, TSMC.
• End customers. Undisclosed AI-inference buyers (neoclouds / model-serving companies / enterprises) holding the $1B in forward contracts.

Chokepoint summary: three of the four scarce nodes in the entire AI supply chain — TSMC leading-edge wafers, CoWoS packaging, HBM3E — are all upstream of Etched and all controlled by suppliers whose priority customer is the company Etched is attacking. The TSMC-linked investment is the one structural mitigant. Names present → this lens is real, not generic.

---

Competitive Advantages (moats)

The moat, if it exists, is specialization-as-performance. By hardwiring attention as fixed-function logic, Sohu eliminates the general-purpose GPU overheads — instruction fetch/decode, kernel-launch latency, memory allocation, thread scheduling — that cause an H100 to run transformer inference at only ~30–40% of peak FLOPS. Etched claims 80–90%+ FLOPS utilization as the source of its throughput edge. If the physical parts validate, that is a genuine architectural advantage a software optimization on a GPU cannot fully close — the overhead is structural to programmability.

Durable moats — graded honestly:

• Scale / process moat: WEAK. Etched has no fab. Its process (N4P) is the same TSMC node incumbents use; it has less buying power. No cost-of-capital or wafer-priority advantage vs NVIDIA.
• Switching-cost moat: DOUBLE-EDGED. Adopting Sohu means rebuilding the serving stack on a proprietary toolchain. Once locked in, that's a retention moat. Before lock-in, it's the single biggest adoption barrier — the migration cost is a tax the incumbent doesn't charge (CUDA is already there).
• IP / architecture moat: NARROW BY CONSTRUCTION. The design is the moat, but it protects only as long as "transformer attention as fixed-function" is both (a) the right thing to hardwire and (b) hard to copy. Taalas is already doing an even-more-extreme version (whole-model mask-ROM); hyperscalers have their own inference ASICs. The IP is defensible; the category is crowded.
• Ecosystem / CUDA moat: THIS IS THE INCUMBENT'S MOAT, NOT ETCHED'S. The deepest moat in the room belongs to NVIDIA. Etched must overcome it, not wield it.

Bargaining power. Over suppliers: near-zero — Etched needs TSMC/SK Hynix far more than they need Etched (partially offset by the TSMC-linked investor). Over customers: entirely dependent on the performance claims surviving independent benchmarking; a 20x tokens/$ advantage would confer real pricing power, an unverified 20x confers none.

The moat's expiry date is written into its foundation. This is the tension the whole thesis rests on, and it collides head-on with Lens 5/13: a moat built on "the algorithm has stabilised" is only as durable as that assumption. See the June-2026 MoE/Mamba-compatibility contradiction flagged at the top — if Sohu is genuinely reconfigurable enough to run MoE and SSM, the moat softens (it's less special) even as the addressable market widens; if it isn't, the moat is real but the TAM is shrinking as the frontier moves to exactly those architectures.

---

Segments

n/a — private, pre-revenue, single-product. There is no segment or geographic revenue to break out: one product (Sohu/FIC), zero recognised revenue, $1B in undisclosed-mix forward bookings. Any "segment" analysis would be fabrication. The only meaningful decomposition available:

• By workload (the relevant axis): Sohu targets transformer LLM inference exclusively. Explicitly out of scope per the architecture: training, diffusion/image/video (Stable Diffusion, Flux), vision-language (LLaVA, Qwen-VL, Llama-Vision), and — per Spheron — MoE routing (DeepSeek V4, Qwen3-235B) and SSM/Mamba scan. Contradicted by Etched's own June-2026 compatibility claims (see top-of-file reconciliation #2).
• By geography: design in the US (San Jose HQ), bring-up/manufacture partly in Taiwan; customers undisclosed.

---

Phase B — Measure performance

Earnings Result → (private overlay) Funding & Traction

No earnings. Substituting the +private Lens-5 swap — funding & valuation trajectory + traction:

Funding trajectory (all, unaudited):

Angel/strategic roster (a real signal): Geoffrey Hinton, Andrej Karpathy, Fei-Fei Li, Peter Thiel, Stanley Druckenmiller, Arthur Mensch (Mistral CEO). The concentration of AI-research names (Hinton, Karpathy, Li, Mensch) is unusually high-signal — these are people paid to know whether the transformer bet is sound. It is also a marketing asset, so weight it, don't worship it.

The tell in the investor base: the Series B is thick with quant/HFT trading firms — Jane Street, Hudson River Trading, Jump, Two Sigma. These are the world's most sophisticated buyers of low-latency inference (they'd be natural Sohu customers) AND the most ruthless underwriters of asymmetric bets. Their presence is simultaneously a customer signal and a "smart money likes the risk/reward" signal.

Traction (unaudited):

• $1B+ in signed forward contracts pre-shipment. Quality/terms undisclosed — treat as a soft booking until deposits/PO terms are shown.
• A0 tape-out returned, claimed first-pass success on TSMC N4P — a genuine de-risking milestone (many chip startups die on a broken first spin).
• First racks ship "summer 2026"; gigawatt-scale by 2027 ambition.
• Preliminary customer tests claimed to validate on Llama, DeepSeek, Qwen, Mamba — company-run, no independent benchmark exists — and this is exactly the claim that collides with reconciliation #2.

Market reaction proxy (private, so read the mark): valuation stepped ~$800M (2024) → ~$1.5B (early 2025) → $5B (Dec 2025), a ~6x in ~18 months. The step-ups are the "market reaction" for a private — and they are pricing near-flawless execution.

---

Earnings Calls → (private overlay) Founder Signal

No earnings calls. Substituting founder interviews/public reasoning (Masters of Scale, Innovators Under 35, BigGo, yespress) [all web, 2026]:

What management is consistently focused on — one message, repeated with discipline:

• "Transformers have stabilised enough to burn into hardware." Uberti's core analogy: x86 won not because it was best but because the ecosystem converged and stayed; he bets transformers follow the same path.
• "ASICs win when algorithms stabilise." The whole thesis compressed.
• "Whoever produces the most tokens will be the most valuable company in the world" — reframing the prize from chips to token-throughput economics.

Tone shift over time (2024 → 2026): The message has hardened, not evolved. In 2024 it was a bold pre-silicon claim; by June 2026 it's "A0 works, racks ship this summer, gigawatt in 2027." The consistency is a strength (conviction, no pivot) and a risk (a single-thesis founder is structurally disincentivised to acknowledge architecture drift). The new thing they started saying in June 2026 — "runs DeepSeek/Qwen/Mamba, supports models of all shapes" — is a softening of the original transformer-purity story and reads as a response to exactly the MoE/SSM critique. That rhetorical move is the most analytically interesting signal in the whole call-substitute (reconciliation #2).

---

Comps → (private overlay) Cap Table + Inference-Accelerator Landscape

Cap-table quality (the +private Lens-7 swap): High-tier and unusually deep. Series B syndicate spans growth (Stripes), foundry-strategic (VentureTech Alliance/TSMC), and elite quant (Jane Street/HRT/Jump/Two Sigma), plus repeat Thiel and an AI-luminary angel bench (Hinton/Karpathy/Li/Mensch). Crossover-fund tell: the classic IPO-proximity signal is a Fidelity/T. Rowe/Coatue entry — not clearly present here (the crossovers are quant trading firms, not mutual-fund crossovers), so on the standard +private readiness rubric Etched looks late-stage growth, not IPO-imminent. n/a — not in private-watch overlay for a formal readiness score; my = readiness 3/5 (growth/late), gated on production shipment.

Peer landscape — inference accelerators (multiples are private → market cap only; specs ):

Comp read: Etched sits at the aggressive-specialization end (Taalas is more extreme; Groq/Cerebras less so). Its $5B mark is the highest of the pure-play inference-ASIC privates except where Cerebras' $1.1B raise implies a comparable or higher valuation — so Etched is priced as the category leader on the strength of claims, not shipped product. Every performance number in this table for Etched is company-sourced and batch-1; n/a — not independently benchmarked is the honest label on the headline.

---

Stock-Price Catalysts → (private overlay) Valuation-Step / Product Catalysts

No public stock. The private analogue — events that re-rated the mark or will re-rate a future one:

• Jun 2024 — Series A + Sohu reveal ($120M; "20x faster than H100" claim) → first big attention spike; validated the idea.
• ~Early 2025 — reported up-round to ~$1.5B (from ~$750M) → thesis gaining, pre-silicon.
• Dec 2025 / Jan 2026 — $500M Series B at $5B → institutional + foundry-linked + quant validation.
• Jun 30 2026 — "stealth exit again": A0 silicon works, $1B contracts, racks this summer → the pivotal de-risking event; moved the story from aspiration to working-silicon-with-bookings. (Sceptic's aside, quoted: "I don't know how you come out of stealth when we've known about your company for 3.5 years".)

Pattern: the mark re-rates on thesis-validation and de-risking milestones (funding, tape-out), NOT on shipped economics — because none exist yet. The next real catalyst is binary and near: either summer-2026 racks ship and an independent party benchmarks batch->256 throughput and $/token, or the timeline slips. That single event is worth more than all prior re-rates combined.

---

Phase C — Judge people & books

Management

• Gavin Uberti (CEO, ~24). Harvard dropout (concurrent BS/MS in CS + Math), Thiel Fellow. Origin: a compiler internship exposed him to low-level hardware; a school project became the company. Track record: this is the track record — no prior shipped chip. Deep conviction, strong first-principles articulation, exceptional at recruiting (400+ senior ex-NVIDIA/TPU/Broadcom engineers and an AI-legend angel bench is itself evidence of magnetism).
• Chris Zhu (co-founder). Harvard (Math + CS), dropped out 2022 with Uberti. Technical co-founder.
• Robert Wachen (President/COO-type). Uberti's college roommate; runs business/ops/BD side.

(1) Track record: unproven at execution — brilliant thesis, zero prior silicon delivered. The A0 first-pass success (if real) is the first hard execution data point and it's a good one. (2) Tenure & skin in the game: founders since 2022, presumably large equity holders (undisclosed) → very high skin in the game; this is a founder-owned single-thesis company. (3) Capital allocation: raised ~$800M and concentrated it into one chip + owned bring-up infra (SJ 2MW DC, Taiwan facility) — coherent and disciplined for the strategy, but all-in with no plan-B silicon. (4) Red flags: the "stealth exit again" and company-controlled benchmarks invite a promotional read; the June-2026 "now runs MoE/Mamba" claim, if it contradicts the physics, is the red flag to watch (over-claiming). No related-party or governance flags surfaced. (5) Archetype: quintessential young founder-visionary (Thiel-fellow, dropout, first-principles, single obsessive bet) — the archetype that either produces a generational outcome or a spectacular zero, rarely the middle. For a hardware company that must out-execute NVIDIA on manufacturing, the lack of a grey-haired operator on the public bench is the relevant gap.

---

Forensic Red Flags

Accounting: n/a — no audited financials, no filings. There is no income statement, balance sheet, or cash-flow statement to forensically examine — the standard revenue-recognition / receivables / SBC analysis is impossible for a private with zero disclosure. The honest forensic points are claim-integrity ones:

• Company-controlled benchmarks, no independent verification. Every headline (500k tok/s, 20x H100, 90% FLOPS util, 10W/chip) is Etched-sourced — the equivalent of unaudited management numbers. 10W/chip in particular is implausible as a full-chip TDP for a reticle-limit die + 144GB HBM3E; it most likely refers to the LVI math-block running at sub-half voltage, not the package — a classic figure-that-flatters if quoted as chip TDP.
• Batch-1 comparison framing. The 500k-tok/s figure is 8 chips at batch size 1 vs GPU figures often quoted at high batch; a fair matched comparison (batch 256) narrows the gap materially — H100 alone hits ~45k tok/s at batch 256. Not fraud — but selective framing that inflates the apparent multiple.
• "$1B in contracts" without disclosed terms — bookings quality unknowable; could include cancellable LOIs.
• The MoE/Mamba compatibility claim vs the transformer-only architecture — reconciliation #2 — is the item a forensic analyst flags hardest: it is either a genuine capability the physics permits or an over-statement, and it is currently unresolvable from public sources.

Regulatory findings (required sub-section):

• SEC (EDGAR LR + AAER): regulatory/regulatory-findings.md (fetched 2026-07-06) — 0 findings; Etched has no CIK and is not an SEC registrant, so no EDGAR enforcement search is possible.
• Non-SEC (FTC/DOJ/FDA/etc.): web search "Etched" (lawsuit OR SEC OR FTC OR investigation OR controversy) 2025 2026 returned no company-specific enforcement, litigation, or investigation hits.
• Item 3 Legal Proceedings: n/a — no 10-K exists (private).
• Conclusion: No material regulatory or legal findings — verified via SEC EDGAR EFTS (LR, AAER, 0 hits), web search (0 hits), and the absence of any filing obligation, as of 2026-07-06. Findings are unaudited per public sources.

---

Phase D — Project & stress-test

Forward Projection → (private overlay) IPO-Readiness & Path-to-Tradeable

No EPS (pre-revenue) → substituting the +private Lens-11 swap. No our model create in --watchlist mode, per skill.

IPO-readiness:

• In private-watch.json? No — Etched is absent from the 28-name overlay, so there is n/a — not in private-watch overlay for the maintained stage/ipo_readiness/catalyst fields, and this run does not write back to that file (wave boundary).
• My readiness = 3/5 (growth/late-stage), NOT IPO-imminent. Rationale: A0 works and there's a $5B mark and elite syndicate, but there is zero shipped product and zero recognised revenue — the gating milestone for any S-1 conversation is production shipment + a real revenue ramp, which is the "summer 2026 → gigawatt 2027" plan, still entirely forward.
• Milestones that unlock a tradeable event, in order: (1) summer-2026 racks actually ship to paying customers; (2) an independent benchmark confirms the throughput/$-per-token claims at realistic batch sizes; (3) forward contracts convert to recognised revenue with disclosed unit economics; (4) a second design win / second-gen part proving it's a roadmap, not a one-shot; (5) evidence the architecture bet survives the MoE/SSM drift. Only after (1)–(3) does an S-1 window (est. 2027–2028 at the earliest ) become credible. A secondary-sale liquidity event or a strategic acquisition (by a hyperscaler or NVIDIA-defensive buyer) is a more probable near-term "tradeable" path than an IPO.

Unit-economics note (Phase-B private add): no published pricing or per-rack cost. TCO vs an H100 at $2.54/hr ($0.141/M tokens) or B200 at $9.36/hr ($0.289/M tokens) is therefore n/a — not calculable until Etched publishes a price. The entire "cheaper than NVIDIA" claim is unpriced.

---

Bull vs Bear

Bull case. The transformer won, and won decisively — it is the substrate under every frontier LLM, and the last two years saw it colonise vision, audio, and code, exactly as Uberti predicted. When an algorithm stabilises, the ASIC beats the GPU on the only axis that matters at scale — tokens per dollar and per watt — because it deletes the 60–70% of GPU silicon burned on flexibility nobody's using for inference. Sohu's A0 came back first-pass on TSMC N4P, $1B is already booked, an AI-Nobel-laureate angel bench and a foundry-linked investor validate both the science and the supply path, and the prize — "the most tokens" — is the single largest market in technology. If the physical racks even half-validate the 20x, Etched is the highest-leverage inference asset outside NVIDIA, and $5B is early. Contrarian view the market is refusing to see: as inference eclipses training as the dominant compute cost, specialization — not generality — becomes the winning strategy, and the incumbents' flexibility becomes a liability they can't shed. "The specialists are winning on inference. And inference is where the money is.".

Bear case (2–3 permanent-impairment risks).

1. The frontier drifts off pure attention faster than a hardwired chip can amortise. The most-deployed open models of 2026 (DeepSeek V4, Qwen3-235B) are MoE; SSM/Mamba and diffusion are rising. A transformer-attention-hardwired die that can't run them (per Spheron) is a shrinking-TAM asset — and a chip needs years of a stable target to earn back an ASIC's NRE. Etched's own June-2026 "we run MoE/Mamba now" claim is either the fix (in which case the moat softens) or spin (in which case the bear case is intact). Either branch hurts the pure thesis.
2. The CUDA/ecosystem moat + migration tax. Adopting Sohu means rebuilding the serving stack on a startup's proprietary toolchain and betting production uptime on it. That is a steep, ongoing tax the incumbent doesn't levy — enough to keep Sohu at the margin of most inference fleets even if the silicon is great.
3. Manufacturing/scale execution against giants. Reaching "gigawatt 2027" means winning TSMC N4P wafers, CoWoS packaging, and HBM3E allocation in competition with NVIDIA/AMD/hyperscalers — the exact suppliers whose priority customer is the target. A 24-year-old first-time-chip team must out-operate the best supply-chain operators on earth. Slip risk is high and structural.

Pre-mortem (it's Jan 2028, the thesis broke — what happened?): Racks shipped late-2026, but the independent batch-256 $/token was ~3–4x better than a GPU, not 20x — good, not category-defining. Meanwhile MoE became the default frontier shape and Sohu's routing support was a slow software emulation, not native; two marquee customers quietly re-weighted to GB200 for fleet flexibility; a follow-on raise came at a flat/down mark; and Taalas/hyperscaler ASICs split the specialist niche. Etched survives as a solid niche vendor — but the $5B was priced for a NVIDIA-killer, and it was a good chip company.

Are the multiples too high? For a private with no revenue, "$5B" is the multiple, and it prices flawless execution + the bet paying off. It is a call option on transformer permanence, fairly struck for a venture bet, richly struck for anyone underwriting the base case.

---

Devil's Advocate (short-seller)

Dismantling the bull case:

• What structurally breaks the money model? The revenue doesn't exist yet, so what breaks is the thesis before it earns: attention stops being the sole thing worth hardwiring. The 2026 evidence is already against purity — the most-used open models are MoE; SSMs and diffusion are climbing. An ASIC needs a multi-year stable target to amortise NRE; the target is visibly moving. A hardwired chip is a bet that the future looks like the past — in the fastest-moving field in technology.
• Where is revenue concentrated / what if it shifts? $1B of undisclosed forward contracts — likely a handful of large, possibly friendly/early customers. If one or two re-weight to GPUs for fleet flexibility (the natural default), the "traction" story deflates. Concentration is almost certainly high and the terms are almost certainly soft (no take-or-pay disclosed).
• Why is the moat weaker than bulls think? (a) It's the same TSMC node as everyone — no process edge. (b) The retention moat (proprietary toolchain) is also the adoption barrier — it cuts against you before it cuts for you. (c) Taalas hardwires whole models at ~10x less power than Etched — a more extreme specialist can undercut on the exact axis Etched competes. (d) The deepest moat in the room (CUDA) belongs to the enemy.
• Most dangerous underestimated competitor? Not NVIDIA (obvious) — the hyperscalers' own inference silicon (Trainium, TPU, Maia, MTIA). They have captive demand, unlimited capital, and no need to sell a chip to anyone — they just deploy it internally and never appear in Etched's contract pipeline. They can absorb the specialization benefit without buying Sohu.
• Worst governance/claim behaviour? Company-controlled benchmarks + batch-1 framing + a "10W/chip" figure that can't be a full-chip TDP + "stealth exit again" theatrics + a June-2026 MoE/Mamba-compatibility claim that appears to contradict the architecture. None are fraud; together they read as aggressive promotion ahead of shipped, verifiable product.
• What must hold for the $5B? That transformers stay the overwhelming inference substrate through ~2028; that Sohu ships on time and independently validates a large (not 2–3x) advantage; that customers eat the migration tax at scale; and that a first-time team wins scarce TSMC/HBM/CoWoS capacity against the biggest buyers alive. All four, simultaneously.
• Valuation if growth disappoints 20–30%? For a pre-revenue $5B mark, "growth disappoints" = timeline slips or the independent benchmark underwhelms → the next round is flat-to-down and the "NVIDIA-killer" premium collapses toward "good niche ASIC," implying a materially lower mark.
• Single scenario that permanently impairs it: a genuinely superior non-transformer architecture (or a transformer++ variant Sohu can't run) becomes the frontier default in 2026–27, before Sohu amortises its NRE — turning a reticle-limit hardwired die into, in the critics' phrase, an "expensive paperweight." Plausibility: not low. The frontier is already partly there (MoE), which is precisely why the plausibility is higher for Etched than for a flexible-GPU vendor.

---

Management Questions (ordered by information value)

1. The MoE/SSM contradiction: you now say Sohu runs DeepSeek, Qwen (MoE) and Mamba (SSM). Architecturally, how does a transformer-attention-hardwired ASIC execute MoE routing and SSM scan ops natively — is this fixed-function, or software emulation on the same fabric, and what's the FLOPS-utilization on those vs on a dense transformer?
2. Independent benchmarks: will you allow a neutral third party (MLPerf or a named lab) to publish Sohu throughput and $/token at batch 32 and batch 256 on Llama-70B, DeepSeek V4, and a diffusion model — and when?
3. Contract quality: of the $1B in forward contracts, how much is take-or-pay with non-refundable deposits vs cancellable LOI, and how concentrated is it across the top 2–3 customers?
4. Full-chip power: what is the package TDP of a Sohu chip (not the LVI math-block figure) and the rack power for the 8-chip FIC, and what is the resulting tokens-per-watt vs GB200?
5. Amortisation math: what is the NRE per Sohu generation, and over how many shipped units / how many months of a stable transformer target must you sell to earn it back?
6. Supply allocation: what contractual TSMC N4P wafer, CoWoS packaging, and HBM3E allocation do you hold for 2026–27, and how does the VentureTech Alliance relationship translate into priority (or not)?
7. Pricing: what is the published $/rack and $/M-tokens, and at what utilisation does a customer's Sohu TCO beat renting a B200?
8. The x86 analogy's failure mode: x86 also got displaced (ARM/RISC-V) once power/efficiency mattered more than the installed base — why doesn't that same dynamic favour a more specialised competitor (Taalas) or a different architecture over Sohu?
9. Roadmap: is there a second-generation part, and does it stay transformer-only or add reconfigurability — and if it adds flexibility, aren't you conceding the pure-ASIC thesis?
10. Migration: what is the realistic engineering cost, in weeks, for a customer to port a production serving stack from vLLM/TensorRT to your toolchain, and who has actually completed it?
11. Hyperscaler threat: how do you win against buyers (AWS/Google/Meta/Microsoft) who can deploy their own inference ASICs internally and will never appear in your sales pipeline?
12. A0 → production: "first-pass success" at A0 — what is the current yield on N4P at reticle-limit die size, and do you need a metal spin or a full A1 before volume?
13. Runway: at current burn (400+ engineers, 2MW DC, Taiwan facility), how many months does the ~$800M raised fund, and does gigawatt-2027 require another large raise — at what implied valuation?
14. Customer diversification: beyond the current pipeline, what fraction of the total inference market is addressable by a transformer-only chip today, and how does that fraction move if MoE/diffusion keep gaining share?
15. The permanent-impairment question: what specific, observable signal would tell you that the transformer bet is failing — and what's the plan-B for the silicon if you saw it?

---