AI Compute Silicon — Cerebras and the Inference Stack (Ai Compute Silicon Sector)

Executive Summary & Action Plan

Verdict

UNDERWEIGHT — no Top Pick. We put fair values on twenty names across the AI compute stack after a $3.3 trillion drawdown, and eighteen of them still trade above ours. The median gap is −39%. Not one name here clears our threshold for new capital, and the two closest to fair value are the category leader and an EDA franchise that already de-rated 42% from its high.

Cerebras (CBRS) remains the anchor and the clearest SELL. The architecture is real, the niche is real, and the economics are not yet in evidence.

Price $176.88 | Market cap $39.4B | Target $86.00 | Upside -51.4% IWANNAVY Fair Value (inherited 2026-07-20) · price as of 2026-07-20 close · all-class market cap basis; 226.533M filing shares × price = $40.1B cross-check · Street $291.09 (n=11, reference only) — the Street capitalizes the $25.0B gross RPO at conversion rates the 36–38% margin guide does not support

Abstract

We mapped the AI compute stack from foundry to frontier lab and underwrote twenty names with SBC-expensed DCFs and normalized-margin multiples, every mark adversarially verified; thirteen of fifteen fresh marks were revised down on review, most often for peak-margin or peak-revenue extrapolation. The sector's July collapse was not a demand event. Every fundamental print inside the drawdown was a record or a raise — TSMC lifted both revenue and capex guidance mid-selloff, Micron posted its first 81%+ gross margin, Samsung printed the largest quarterly operating profit in technology history — while the de-rating ran on a hawkish Fed debut, a positioning unwind, and an efficiency shock from Moonshot's Kimi K3. What the drawdown did not do is create value: the complex entered it priced for perfection and remains above disciplined fair value almost everywhere. Rent is migrating down to the foundry and out to the buyers as cost per token falls faster than accelerator revenue rises. Principal risks are capex durability, memory normalization, and a Cerebras lockup worth 1.9x its float.

Forecast Path

The path prices continued multiple compression against intact near-term earnings, not a demand break. Bands measure dispersion, not scenario endpoints.

Scenario probabilities are judgmental, not calibrated. The dispersion is genuine: the SOX sits 0.3% above its bear-market threshold, so the next two weeks of prints settle a question the tape has not.

Action Plan

Own nothing here on valuation alone. Reduce the extremes — Arm (ARM), CoreWeave (CRWV), Intel (INTC), Micron (MU) and Astera Labs (ALAB) all trade at more than double our marks. Hold quality at fair value: NVIDIA (NVDA) and Synopsys (SNPS) are the only two names where price and value have met. Revisit the enabler layer — Broadcom (AVGO), Cadence (CDNS), TSMC (TSM) — on a 20–25% drawdown, since those are the franchises we would want to own at a price. For the anchor, the ladder below is unchanged from our 2026-07-20 company work.

• Prob-weighted 12M index return −12.4% | P(sector thesis wrong) 35%, judgmental | Confidence Medium-High on the valuation framework, Medium on cycle timing

---

What Happened

The consensus account of this drawdown is wrong on sequence, and the sequence matters.

Broadcom (AVGO) guided Q3 AI semiconductor revenue to ~$16.0B against ~$17.2B expected on 2026-06-03, and the complex lost roughly $1.4T that week. It then fully recovered, and the SOX made an all-time high near 14,655 on 2026-06-22. The $3.3T is measured from that peak. The trigger on 2026-06-23 was new Fed Chair Kevin Warsh's hawkish debut — rates held, but the dot plot flipped to show a 2026 hike against inflation running 3.6–4.1% (CNBC, 2026-06-23). Anyone reading the selloff as a verdict on Broadcom's guide has the causality backwards.

The acute leg came on 2026-07-17, when Moonshot AI released Kimi K3 at WAIC Shanghai — a 2.8T-parameter open-weight model with a 1M-token context at roughly half the cost of Claude Opus 4.8. The SOX fell 5.7% intraday and Micron lost 13%, about $138B in a session (Reuters, 2026-07-17). This is an efficiency shock, not a demand shock, and the two falsify on entirely different evidence.

The fundamental content of the drawdown is remarkably thin. TSMC (TSM) raised FY26 revenue growth to >40% and capex to $60–64B during the selloff, plus $100B for four Arizona fabs, calling the buildout "durable, not a short-cycle project" (TSMC Q2'26). Micron (MU) posted record $41.46B revenue at its first-ever 81%+ gross margin. Samsung (005930.KS) printed KRW 89.4T of quarterly operating profit — the largest in technology history — and fell anyway. Hyperscaler capex is still being raised: Amazon ~$200B, Microsoft ~$190B, Alphabet $180–190B, Meta $125–145B, plus Oracle at $55.7B FY26 and ~$70B FY27, aggregating to roughly $750–780B for 2026 against ~$410B in 2025.

Korea's 2026-07-13 crash (KOSPI −8.95%, SK Hynix −15.4%, its worst day on record) was plumbing rather than fundamentals: the Seoul line fell while its newly listed Nasdaq ADR rose ~13% the same day, a listing-venue dislocation compounded by leveraged-ETF deleveraging after SK Hynix raised $26.5B in the largest US listing ever by a foreign company.

Two sessions of rebound leave the SOX at 11,743.85, just 0.3% above the 11,708 bear-market line. The question is unresolved and the earnings gauntlet settles it.

---

Where Cerebras Actually Sits

Inference splits into two different physics problems, and almost every architecture here is a bet on one side of that split. Prefill — digesting the prompt — is compute-bound, and GPUs are the cheapest FLOPs available. Decode — emitting tokens one at a time — is bandwidth- and latency-bound, because producing a single token requires streaming every weight in the model past the arithmetic units. Reasoning models detonated decode volume, since chain-of-thought generates enormous numbers of intermediate tokens the user waits for but never sees.

Wafer-scale removes three things from the decode loop: the off-package trip to HBM (weights sit in on-wafer SRAM), the inter-chip all-reduce that costs a synchronization per layer per token, and the need to batch users into a queue to reclaim efficiency. That is why OpenAI committed 750MW while still buying GPUs — it runs both halves of the workload.

The niche is real and independently validated. SambaNova and Intel shipped the split as a product ("GPUs for prefill, RDUs for decode"). Google forked its TPU line for the first time in over a decade, tripling per-core SRAM on the inference part and adding a collectives engine with 5x lower latency — the same architectural conclusion, reached by the most experienced ASIC team in the industry. And NVIDIA paid to buy into it, sizing low-latency inference at roughly 25% of AI cluster compute.

But four limits bound the thesis, and one of them is specific and damaging. SRAM density has largely stalled at recent nodes, so capacity — not compute — is the binding constraint. The Next Platform's assessment is blunt: the compute-to-SRAM ratio on WSE-3 is wrong for low-latency inference. Cerebras had to gang three, then four CS-3 machines together not for the compute, but because that was the only way to hold the model weights near it — then stopped disclosing system counts in benchmark results (The Next Platform, 2026-05-15). Sparse MoE attacks the thesis directly by shrinking per-token bandwidth demand, which is the exact pain wafer-scale exists to solve. Dedicated silicon is stranded capital when a model cools. And a genuine architectural moat should surface as pricing power; so far it surfaces as volume at deteriorating margin.

Two structural facts cut the other way and are the strongest available bull rebuttal. Cerebras carries no HBM stacks, so it is structurally unconstrained by CoWoS — the single tightest chokepoint in AI compute — and can scale toward 750MW without queuing behind NVIDIA at TSMC's packaging lines. Conversely, it is a conventional DRAM buyer through external weight-streaming memory, which means the contract-price spike that produced SK Hynix's record margins is a direct Cerebras cost headwind, and a plausibly under-explained driver of the core gross margin guide-down from 46.5% to 36–38%.

---

The Bottleneck Ladder

What gates volume and what captures rent are different lists, and conflating them is the common error.

TSMC is the standout and deserves emphasis: Q2'26 revenue $40.2B (+33.7%), gross margin 67.7%, net income $22.37B (+74.7%), with AI silicon around $13.31B — roughly a third of total revenue. It is architecture-agnostic. GPU, wafer-scale, LPU, TPU, MTIA — nearly all of it queues at the same fab, and TSMC clips ~68% gross margin regardless of which bet wins. It is the only position on this map that does not require picking a winner, which is precisely why we would want to own it at a price and do not at 20.4x.

The decisive economic point: substitution compresses margins across the entire accelerator layer while the foundry's take rate holds. Cerebras guided margins down, Broadcom's AI outlook disappointed, HBM prices softened — and TSMC printed 67.7% and raised capex. Rent is migrating down to the foundry and out to the buyers, who capture the savings as lower cost per token. Cost per token is falling faster than accelerator-vendor revenue is rising. That is excellent for model companies and for TSMC, ambiguous for NVIDIA, and difficult for sub-scale challengers funding custom silicon against falling ASPs.

---

The Memory Layer Is Mis-narrated

This is the most misunderstood line in the sector, and it changes what the record prints mean.

HBM wafer revenue was overtaken by DDR5 64GB RDIMM modules in Q1 2026, and HBM profitability fell below it in the same quarter (TrendForce, 2026-06-02). The mechanism is that HBM prices annually, so it could not participate in a quarterly supply squeeze, while conventional DRAM repriced and captured all of it. SK Hynix's Q1'26 DRAM ASP rose in the mid-60% range QoQ on flat bit growth — a mix cycle, not a volume cycle.

So the profit engine behind every record memory print is a conventional-DRAM supply squeeze caused by HBM capacity conversion — HBM is driving memory profits indirectly, by cannibalizing its own substrate. That mechanism reverses when HBM4 capacity lands. Do not underwrite 71.5% operating margins as HBM-driven.

One correction to our own prior work: the "HBM4 slipped to 2027" framing we carried on 2026-07-13 is now outdated. Micron confirmed on 2026-06-24 that HBM4 is in high-volume shipments for its lead customer's platform, and Jensen Huang confirmed all three suppliers qualified and in production for Vera Rubin. It is HBM4E that sits in 2027. Separately, 2027 HBM contract negotiations are reported pricing up, not down — the softness is a 2026 phenomenon locked in by annual contracts.

Micron's mark moved the furthest of any name here as a result. Our prior $1,300 (2026-06-30) predated the peak-margin discipline we applied to the Korean names and would have implied 50% upside purely from a methodology mismatch. Re-underwritten to $300.

---

What to Watch Now

• The capex verdict, 7/22–7/30: Alphabet (7/22), Microsoft and Meta (7/29), Amazon (7/30). Nothing in published guidance supports a demand-destruction thesis; the de-rating priced a hedge that has not appeared. Any single hyperscaler guiding 2027 capex flat or down — especially citing efficiency — converts a multiple event into an earnings event.
• SK Hynix (7/29) and Samsung (7/30), both confirmed from company IR. Watch the conventional-DRAM versus HBM ASP split, not the headline. This is where the mix-cycle thesis is settled.
• TSMC monthly revenue, ~the 10th of each month — the shortest-latency read in the entire stack, leading every downstream P&L by one to two quarters. Two consecutive months of decelerating YoY with the three-month rolling below +20% is the change-of-view threshold.
• Cerebras Q2 (mid-Aug to early Sep, estimated): the FY26 guide of $855–865M against $385.3M of H1 actual-plus-guide requires H2 to deliver $470–480M — two consecutive quarters of +11–16% QoQ growth immediately after a +1.4% QoQ quarter. The gross margin print is the event, not revenue.
• NVIDIA (8/26) and Broadcom (9/3): NVIDIA is roughly flat through the entire de-rating and carries the sector's remaining leadership. Broadcom revisits the guide that started the June air pocket.

What to Watch Next (Technology)

• Prefill/decode disaggregation as the default rack topology — the open question is no longer whether racks go heterogeneous but whether the decode tier is merchant-supplied or NVIDIA-internal.
• Sub-8-bit numerics — Google's TPU 8 added FP4, and half of Meta's 293x throughput gain came from adopting 4-bit MX4; format adoption is currently buying more performance than process migration.
• Memory-tier architecture: SRAM versus HBM4/HBM4E versus 3D DRAM — this decides whether the latency niche survives sparse MoE.
• Power- and thermal-first design, where performance per watt is the currency and the grid, not the fab, limits 2027 deployments.
• Whether the scale-up memory domain stays proprietary — Google's collectives engine, Meta's 72-unit shared domain and Etched's cluster-scale memory all attack the bottleneck NVLink defends.

---

Fair Value Notes

All twenty marks are log-medians of a strict SBC-expensed DCF and two normalized forward cash-flow multiples, priced off the frozen 2026-07-20 pack. Thirteen of fifteen fresh marks were cut on adversarial review; the recurring failure was normalizing the margin and then re-importing the cycle premium through the multiple.

At fair value (2). NVIDIA (NVDA) $210 HOLD — strict DCF $133, EV/FCF $210, EV/EBITDA $249; the leader trades below its own #2 on forward EBITDA, which is the map's central pricing anomaly (inherited 2026-07-15). Synopsys (SNPS) $384 HOLD — DCF $326, 20x normalized FY27E EBITDA $446, 30x FY27E FCF $384; nothing broke in the P&L, the multiple broke, hit by a 2026-07-17 demonstration of an autonomous 45nm tapeout on open-source EDA tools.

Enablers we would own at a price (4). Broadcom (AVGO) $320 HOLD — the best-executed custom-silicon franchise, priced for its own guidance to be a floor. Cadence (CDNS) $278 HOLD — the best business in the coverage universe at the wrong price; the verifier overturned an initial SELL as inconsistent with the house rating ladder at only −15.7%. TSMC (TSM) $284 HOLD — owns the real bottleneck and is the cheapest primary-metric name here, but a record 67.7% margin plus a $60–64B capex year is a lot to pay forward; the initial mark carried a terminal-discounting error worth 14%. AMD $355 HOLD — strict DCF $188 against a FY27E EV/EBITDA above NVIDIA's, indefensible for the #2 (inherited 2026-07-15).

Priced beyond the evidence (6). Qualcomm (QCOM) $130 HOLD — half genuine value, half melting ice cube, with $5.7–7.8B of Apple modem revenue going to zero by FY28. Arista (ANET) $128 SELL — winning Ethernet-versus-InfiniBand and losing Ethernet-versus-NVIDIA simultaneously. Vertiv (VRT) $186 SELL — the power-and-thermal bottleneck is fully in the price; the 23.3% margin being capitalized is scarcity rent from a business that earned 8% in 2019. Samsung (005930.KS) ₩150,000 and SK Hynix (000660.KS) ₩1,060,000, both SELL — peak-margin normalized (inherited 2026-07-13). Nebius (NBIS) $105 SELL — the initial mark counted one multiple twice and discarded the DCF entirely.

Extremes (8). Credo (CRDO) $113 SELL — 68% gross margins with three accounts at 77% of revenue. Marvell (MRVL) $96 SELL — the genuine #2 custom-ASIC franchise, priced at parity with the leader. Cerebras (CBRS) $86 SELL — DCF $86, 30x normalized FY28E EBITDA $95, 12x FY27E sales $83; negative forward FCF makes EV/FCF unusable (inherited 2026-07-20). Astera Labs (ALAB) $116 SELL — an excellent business at 222x trailing EBITDA, requiring it to become a $25–30B revenue company by 2033. Micron (MU) $300 SELL — re-underwritten; an 86% gross margin is not a run rate, and essentially all revenue above ~$14B/quarter is price, not volume. Intel (INTC) $33 SELL — a 278% H1 re-rating prices a sovereign-backed foundry turnaround with zero committed external 14A customers and negative FCF. CoreWeave (CRWV) $24 SELL — a 45%-levered GPU lessor whose 56% adjusted EBITDA margin disguises depreciation at ~55% of revenue and a near-zero pre-tax ROIC against an 11.5% cost of capital. Arm (ARM) $58 SELL — at 267x EV/EBITDA, the price requires roughly $38B of FY2033 free cash flow from a company that generated $979M in FY2026.

---

The Wider Universe

The user asked for breadth, and the covered twenty are not the whole map. The highest-relevance names absent from our coverage, ranked by exposure to the Cerebras thesis specifically: Disco (6146.T) — near-monopoly in wafer grinding and dicing, and Cerebras consumes a full 300mm wafer per part, making it the most wafer-levered product in AI compute. Zhongji Innolight (300308.SZ) — the world's largest 800G/1.6T transceiver maker at roughly $178B, essentially invisible to US-only screens, and Cerebras runs its fabric over standard Ethernet rather than proprietary interconnect. Eaton (ETN) and GE Vernova (GEV) — every megawatt of the OpenAI staging passes through electrical gear and grid capacity, and turbine slots are sold out for years. Advantest (6857.T) — near-monopoly in HBM and high-end SoC test, where test time scales superlinearly with die area. Nanya Technology (2408.TW) — pure conventional-DRAM leverage with no HBM to dilute the signal, which is the cleanest listed expression of the mix-cycle insight above. Also material: Comfort Systems (FIX), Monolithic Power (MPWR), Rambus (RMBS) as a royalty on HBM volume immune to HBM price, Ajinomoto (2802.T) and Ibiden (4062.T) in ABF substrates, SanDisk (SNDK) and Seagate (STX) in the storage leg, and ASPEED (5274.TWO) with ~70% share of server BMC.

Private names that matter: SambaNova at $11B (first close of a $1B Series F led by General Atlantic, 2026-07-08, with JPMorganChase as anchor customer), Etched, d-Matrix (Corsair in full production since June 2026), Positron, Tenstorrent, and G42/Core42, which is simultaneously Cerebras's anchor customer, operator and shareholder — the single largest concentration in the story.

One correction to our prior sector work: Meta "Iris" does not appear to exist. Meta's accelerator line is MTIA 100–500 with Broadcom as design partner, and we could not source an "Iris" part. Our 2026-07-12 AI Value Chain report carried that name; treat it as retired pending a citation.

---

Catalysts & Risks

The catalysts below are dated and falsifiable; the risks are ranked by impact on sector fair value over twelve months.

Falsification. The bearish stance here is wrong if TSMC delivers Q3 at or above $45.8B and guides 2027 capex up while monthly revenue holds 3-month rolling YoY at 25%+, and Micron guides FQ1-27 flat-to-up on both revenue and margin — that removes the capex-durability and memory-peak legs simultaneously. It is also wrong if served-token volume outruns the decline in price per token for two consecutive months post-Kimi K3, which would make the efficiency shock accretive to inference specialists rather than deflationary. The bullish stance is wrong if any named hyperscaler guides 2027 capex flat or down citing efficiency, or if Broadcom prints AI semis below $15.5B with a Q4 guide at or under $16.0B.

---

References

• TSMC Q2'26 results · TSMC raises 2026 guidance (Yahoo, 2026-07-16) · AI chips drive a third of TSMC revenue (The Next Platform, 2026-07-16)
• Micron FQ3'26 record results (2026-06-24) · Samsung Q2'26 preliminary (CNBC, 2026-07-07) · SK hynix Q1'26 results (2026-04-23)
• HBM overtaken by DDR5 64GB on revenue and profitability (TrendForce, 2026-06-02) · Server DRAM +13–18% QoQ in Q3'26 (TrendForce)
• Moonshot Kimi K3 (Reuters, 2026-07-17) · Tech selloff and Warsh's debut (CNBC, 2026-06-23) · The $3.3T chip selloff (Yahoo, 2026-07-17)
• NVIDIA–Groq non-exclusive licensing agreement (Groq newsroom, 2025-12-24) · Deal structure (Constellation Research) · NVIDIA FY26 10-K, Note 2
• Cerebras 424B4 prospectus (2026-05-14) · Cerebras Q1'26 10-Q · Wafer-scale SRAM assessment (The Next Platform, 2026-05-15)
• Google TPU 8 fork (The Next Platform, 2026-04-24) · Meta MTIA roadmap (The Next Platform, 2026-04-08) · SambaNova $11B first close (2026-07-08)
• Broadcom Q2 FY26 and the $16.0B AI guide (2026-06-03) · AI demand begins to justify buildout cost (Bloomberg, 2026-06-25) · Bain: $2T revenue needed by 2030
• Companion house reports: Cerebras company tearsheet (2026-07-20) · Korea AI memory complex (2026-07-13) · AI value chain software (2026-07-12)