Groq

Phase A — Understand the business

Lens 1 · Company Overview

Groq is a US AI-inference company founded in 2016 by Jonathan Ross (creator of Google's original TPU) and Douglas Wightman, both ex-Google-X. Its differentiator is a custom silicon design — the LPU (Language Processing Unit) — purpose-built to run already-trained language models at very low latency, not to train them.

The business has run on (historically) two and now arguably four revenue lines:

1. GroqCloud — a fully-managed, OpenAI-API-compatible inference cloud ("tokens-as-a-service"); the developer swaps three lines of code (key, base URL, model name) to switch from OpenAI. This is now the company's core go-forward business.
2. Hardware / on-prem systems — GroqRack appliances (64–576+ LPU chips per rack) sold to large buyers, most notably Aramco Digital in Saudi Arabia.
3. Enterprise / sovereign contracts — large bespoke deployments (the Saudi/HUMAIN build).
4. Technology licensing — new, one-off: the Nvidia license (see Lens 5).

The defining structural fact of this company as of mid-2026: On 24 Dec 2025, Nvidia signed a non-exclusive, perpetual licence to Groq's entire inference IP/patent portfolio + software stack, headlined at ~$20B and structured as ~$17B in cash across three installments by end-2026 — and ~90% of Groq's staff, including founder/CEO Jonathan Ross and president Sunny Madra, moved to Nvidia. Groq says it remains independent under new CEO Simon Edwards (ex-CFO, elevated Apr 2026), and GroqCloud "continues without interruption". So the entity analyzed here is the post-licence Groq: it kept the customers, the cloud, the brand, and a war-chest — but Nvidia now owns a perpetual right to the same architecture, and the people who built it. Every lens below must be read through that lens.

Customers/partners (research-layer customers.csv is empty — all ``): Aramco Digital / HUMAIN (sovereign Saudi build, the anchor), Samsung (foundry + Series-D investor), Cisco, and ~2M registered developers on GroqCloud (up from ~356k in 2024).

Lens 2 · Supply Chain

Upstream → Groq → end customer, named stakeholders (, no supply-chain.csv):

• Foundry (the chokepoint). Current-gen LPU is fabbed by GlobalFoundries on a 14nm node — a deliberately mature, cheap, HBM-free process. Next-gen ("GroqChip 4nm", ~500MB SRAM, ~98B transistors, ~150 TB/s) is fabbed by Samsung Foundry on SF4X (4nm), mass-production targeted 2H25. Note: SemiAnalysis's 2024 piece cites Marvell as a margin-taking partner; later 2025–26 sources name Samsung as the foundry — treat the Marvell reference as stale.
• Memory. Architecturally, Groq's headline input-cost advantage is what it doesn't buy: no HBM, no CoWoS advanced packaging — model weights live in on-chip SRAM (230MB current / ~500MB next-gen). This sidesteps the single most contested supply bottleneck in the whole AI complex (HBM/CoWoS), a genuine structural positive.
• System integration. A production model is sharded across hundreds of chips — e.g. 576 chips for Mixtral 8x7B, requiring 144 CPUs and 144TB of RAM as host overhead. So the "supply chain" per deployed model includes a large conventional server BOM, not just LPUs.
• Downstream / end customer. Concentrated: Aramco Digital → HUMAIN in Dammam (EMEA's largest AI compute centre, ~19,000 LPUs initially) is the marquee deployment. Plus the long tail of GroqCloud developers.

Single-source dependency: next-gen volume rides on Samsung SF4X yield (reported breaking 80% in Apr 2026). And the demand side proved single-source-fragile — see Lens 5.

Lens 3 · Competitive Advantages (moats)

positioning.md/bottlenecks.md exist for the AI topic but Groq isn't in them; this is ``-grounded.

The real moat (latency). On a like-for-like single-stream basis Groq is genuinely class-leading: ~800 tok/s on Llama-3 70B vs ~90 tok/s on an H100 (~9–15×), with deterministic, sub-10ms first-token latency because there's no off-chip HBM round-trip. Determinism (predictable per-token latency) is a real engineering edge for real-time/agentic workloads.

Why the moat is shallower than it looks:

1. The IP moat is now legally non-exclusive — Nvidia bought a perpetual licence to it. The single best defence Groq had (its patent estate) is now also held by the dominant incumbent. A moat you've licensed to the castle next door is not a moat.
2. The people moat left. ~90% of engineering, plus Ross, went to Nvidia. The tacit knowledge that compounds a silicon roadmap walked out the door.
3. Switching cost is deliberately near-zero — by Groq's own marketing. "Change three lines of code" cuts both ways: customers can leave as easily as they arrived.
4. Architectural ceiling. SRAM-only means large models / high-batch throughput fan out across hundreds of chips; Groq is "not competitive architecturally at all for throughput-optimized scenarios". The moat is a latency niche, not the inference market.

Bargaining power: weak over its anchor customer (KSA walked to Nvidia/AMD — Lens 5), weak over Nvidia (its licensee and its most dangerous competitor), reasonable over Samsung (a foundry that invested in it and needs the volume).

Lens 4 · Segments

segments.csv is header-only — no `` segment data exists. Public reporting does not break Groq into audited product/geographic segments. Qualitatively :

• By product: shifting hard from hardware/systems sales (the Saudi GroqRack model) toward GroqCloud tokens-as-a-service — new CEO Edwards is explicitly betting the cloud TAM > the chip-design TAM. Plus a one-off licensing windfall (Nvidia) that is not recurring.
• By geography: historically dangerously concentrated in Saudi Arabia (the $1.5B Aramco/HUMAIN expansion); GroqCloud's developer base is global but monetization is thin.

n/a for any segment revenue/EBITDA split. The honest read: there is no clean segment P&L for a private company that just had its cap structure and headcount restructured by the Nvidia deal.

---

Phase B — Measure performance

Lens 5 · Traction & unit economics (+private swap of "Earnings Result")

The traction story is one of a spectacular guide-down, then a pivot funded by a windfall.

Why the guide-down (this is the single most important performance fact): the $1.5B Saudi supply deal slipped, and after signing, Saudi Arabia began importing AI chips from Nvidia and AMD instead. The anchor customer that justified the whole 2025 hardware ramp diversified away from Groq. A 75% revenue miss against management's own number is a severe credibility event.

The offsetting windfall: the Nvidia licence delivered ~$17B cash (3 installments through end-2026); $7.6B was distributed to shareholders in Feb 2026 (~$64/share, ~¾ of ownership). So existing holders have already been substantially paid out — a crucial structural fact for any new private-market buyer (you are buying the stub after the cash-out).

Go-forward funding: Groq is raising ~$650M pro-rata from existing backers (Disruptive, Infinitum backstopping) to fund GroqCloud capacity + next-gen LPU. It had separately sought ~$500M specifically to fulfil Saudi orders.

Unit economics — the unresolved question. SemiAnalysis: GroqCloud's $0.27 / million-token pricing may not reflect genuine unit economics given the 576-chip + 144-CPU + 144TB-RAM system overhead per model; "not competitive architecturally at all for throughput-optimized scenarios," with fundraising-driven pricing a live possibility. Gross margin: n/a (no public/audited figure). Burn: n/a — not disclosed, but the simultaneous $650M and $500M raises against a $500M revenue year signal heavy cash consumption.

Lens 6 · Founder/exec sentiment trend (+private swap of "Earnings Calls")

No earnings transcripts (transcripts/ empty). Sentiment via public posture,:

• Pre-deal (2024–mid-2025): maximalist — Ross's "double the world's AI compute," $2B revenue guide, "50% of global inference via GroqCloud" ambition.
• The inflection (Jul 2025): the $1.5B guide-cut — the first public admission the demand wasn't there.
• Post-deal (Dec 2025 →): the messaging flips from "we will displace Nvidia" to "we licensed Nvidia and will keep operating GroqCloud independently". New CEO Edwards (a finance operator, not the visionary founder) reframes the story as a focused cloud business, not a frontier-silicon insurgency.
• What they stopped saying: the "Nvidia-killer" framing. The recurring new phrase is "independent... GroqCloud continues without interruption" — defensive, reassurance-led tone.

Lens 7 · Cap table & secondary marks (+private swap of "Comps")

Cap table / syndicate quality — genuinely tier-1, an IPO-proximity tell historically:

Crossover/mutual-fund presence (BlackRock, Neuberger, a West-Coast mutual-fund manager) is normally an IPO-proximity signal — but read it against the secondary marks below.

Secondary marks — a sharp markdown (the live valuation tell):

• Last primary round: $6.9B (Sep 2025).
• Forge: ~$29.77/share (7 May 2026); Nasdaq Private Market: ~$34.74/share (1 Jun 2026).
• IPO Club initiated coverage at a $2.8B valuation.

The $2.8B secondary-context vs $6.9B primary is roughly a ~60% markdown — consistent with the company having (a) distributed $7.6B out to holders (so the equity now owns a smaller residual), and (b) lost its IP-exclusivity and its founding team. This is the single cleanest market signal that the stub is worth far less than the pre-deal headline.

Peer set (by mechanism/inference-silicon, not P/E — there's no EPS): Cerebras (wafer-scale, single-chip huge memory), SambaNova (reconfigurable dataflow), Etched (transformer-only ASIC Sohu), d-Matrix (Microsoft-backed; raised $275M @ $2B), plus the in-house giants Groq can't out-scale: Nvidia (~80–85% of DC AI accelerators), Google TPU v6, AWS Trainium/Inferentia, AMD MI3xx. The independent-startup cohort combined is ~1% of the market.

Lens 8 · Catalysts that moved the mark (+private swap — funding/product events)

Events that materially re-rated the private mark or the narrative :

• Sep 2024 — Aramco/HUMAIN partnership announced (world's-largest inferencing DC) → demand-story peak.
• Dec 2024 — first KSA cluster deployed.
• Feb 2025 — $1.5B Saudi expansion (Dammam) announced → valuation tailwind into the Series E.
• 17 Sep 2025 — $750M Series E @ $6.9B.
• 31 Jul 2025 — 2025 revenue guide cut $2B→$500M (Saudi slip; KSA buys Nvidia/AMD) → first negative re-rate.
• 24 Dec 2025 — Nvidia ~$20B / ~$17B-cash non-exclusive licence + acqui-hire → the regime change.
• Feb 2026 — $7.6B shareholder distribution (~$64/sh).
• Mar 2026 — Senate (Warren/Blumenthal) + DOJ/FTC scrutiny of the deal.
• May–Jun 2026 — secondary marks ~$30–35/sh; $650M pro-rata raise.

Pattern: the mark is driven almost entirely by (a) the Saudi sovereign-demand story and (b) the Nvidia relationship — i.e. two binary, exogenous, counterparty-controlled events, not by organic GroqCloud monetization. That concentration is itself the risk.

---

Phase C — Judge people & books

Lens 9 · Management

• Jonathan Ross (founder, ex-CEO) — now at Nvidia. A-grade builder: started Google's TPU as a 20% project, designed core elements incl. the systolic-array approach; TPU eventually powered >50% of Google compute. The single most important management fact is that he and ~90% of the team are gone to Nvidia. Founder/visionary archetype — and the visionary is no longer running it.
• Simon Edwards (CEO, Apr 2026) — ex-CFO. A finance operator elevated to steward the cloud business and the cash post-deal. Sensible for a "manage the war-chest, narrow the focus" phase; not the profile that wins a silicon-architecture war against Nvidia. Track record as CEO: unproven (weeks in role at analysis date).
• Capital allocation: the defining decision — licence the crown-jewel IP to Nvidia for ~$17B cash and let the team go rather than keep fighting. Defensible as risk-adjusted value capture (the bear case below says the standalone path was deteriorating); but it is, functionally, a partial liquidation of the company's reason to exist. Then distributing $7.6B to holders (not retaining it all to fund the cloud) is a shareholder-friendly but growth-cautious signal.
• Skin in the game / insider ownership: n/a — private, not disclosed (insider-transactions.csv absent). The $7.6B distribution implies founders/early holders have already monetized substantially.
• Founder vs professional manager: transitioned, in one step, from founder-led insurgent to professional-manager-run cash-rich niche cloud. That transition is the story.

Lens 10 · Forensic Red Flags

No filings to forensically read (private, no CIK) — financials.csv empty. Red flags are therefore structural/qualitative, all ``, unaudited:

1. Guidance integrity — severe. A 75% revenue guide-cut ($2B→$500M) inside one year is the kind of miss that, in a public company, triggers restat's-worth of scrutiny. It says management's forecasting (or its dependence on a single sovereign counterparty) was badly miscalibrated.
2. Revenue-quality opacity. The go-forward number will be muddied by a one-off $17B licence booked alongside genuine recurring GroqCloud revenue. Watch for "revenue" figures that fold in licensing — that is not a recurring run-rate. `` reporting already lists licensing as a "fourth revenue stream," which risks flattering the growth story.
3. Pricing vs cost. $0.27/Mtok against a 576-chip/144-CPU system overhead → SemiAnalysis's open question of subsidized pricing. Unaudited gross margin = a black box.
4. Cap-structure churn. A $7.6B distribution + simultaneous $650M and $500M raises in ~6 months is a lot of balance-sheet motion; the residual equity's economics are hard for an outside buyer to model.

Regulatory findings (required sub-section). Per companies/groq/regulatory/regulatory-findings.md (fetched 2026-06-18): Groq has no CIK — no SEC EDGAR enforcement (LR/AAER) is possible; total_sec_findings = 0. No 10-K Item 3 exists (private). Non-SEC web search result:

• The live regulatory exposure is on the Nvidia deal, not on Groq's accounting. Senators Warren and Blumenthal opened an inquiry (~20 Mar 2026, Apr 3 response deadline) and urged DOJ/FTC to investigate whether the ~$20B "non-exclusive licence + acqui-hire of ~90% of staff" is a disguised acquisition structured to dodge antitrust (HSR) review. This is material: an adverse DOJ/FTC outcome could unwind or claw at the deal economics that now underpin Groq's balance sheet.
• No FTC/DOJ/FDA/CFPB consent decree, settlement, fine or penalty against Groq itself surfaced in web search as of 2026-06-18.
• Net: No material regulatory/legal findings against Groq as an operating entity (verified via SEC EDGAR EFTS [no CIK], web search, and the absence of any 10-K). The one real regulatory variable is the antitrust review of its counterparty deal with Nvidia.

---

Phase D — Project & stress-test

Lens 11 · IPO-readiness & path-to-tradeable (+private swap of "Forward Projection")

research/private-watch.json lists groq but with no structured stage/readiness/catalyst object, so this is web-only.

• Stage (on the 1–5 scale, 4 = pre-IPO/secondary-active): ~3.5–4. Secondaries are actively trading (Forge/NPM marks exist), crossover investors are on the cap table — but the trajectory is away from an IPO, not toward one.
• IPO readiness — deteriorated. Multiple sources read the post-Nvidia entity as less likely to IPO: the deal "carved out Groq's core technology and much of its talent, leaving the remaining company centered primarily on its LPU cloud business". IPO Club's nominal "mid-to-late 2026 window" predates/ignores the Nvidia deal and looks stale.
• Milestones that would unlock an S-1: (1) demonstrate organic GroqCloud ARR growth with disclosed positive unit economics (not licence-inflated); (2) re-establish a large recurring customer to replace the Saudi hardware revenue; (3) clear the DOJ/FTC overhang on the Nvidia deal; (4) ship the Samsung 4nm next-gen LPU at scale to prove the roadmap survives the brain-drain.
• Estimated window: `` no credible IPO within 12–24 months — the more likely tradeable outcome is continued secondary trading at a marked-down (~$2.8–3B context) valuation, or eventual absorption/wind-down if GroqCloud doesn't reach standalone economics. Show-the-work: $6.9B primary (Sep'25) → ~$2.8B secondary-context (2026) ≈ −60%, after a $7.6B cash-out to holders → the residual is a cash-cushioned but strategically-diminished stub.

(No Brier forecast logged — --watchlist unattended; the resolvable binary worth tracking later is "Groq files an S-1 by 2027-12-31", p≈0.20.)

Lens 12 · Bull vs Bear

Bull case. Inference is the larger, faster-growing half of AI compute ("80%+ of demand," growing "a billion times" per Huang). Groq owns a real latency/determinism edge for real-time and agentic workloads, an HBM/CoWoS-free cost structure that dodges the worst supply bottleneck, ~2M developers, a credible 4nm Samsung roadmap, and — uniquely among startups — a fortress balance sheet from the ~$17B Nvidia cash. A focused, well-capitalized "fastest-inference neocloud" could compound in the latency niche even at ~1% share, and the Nvidia validation de-risks the technology.

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

1. The moat is licensed to the enemy and the builders are gone. Nvidia holds a perpetual licence to the same architecture and absorbed ~90% of the team incl. Ross. Groq must now out-execute Nvidia at Nvidia's own newly-licensed game, with a finance CEO and a depleted bench.
2. Demand concentration already broke once. The Saudi anchor diversified to Nvidia/AMD, forcing a 75% guide-cut. There is no proven, large, recurring demand base to replace it.
3. Token economics are compressing under it. DeepSeek's V4 cut inference prices ~75%; SemiAnalysis already questioned whether $0.27/Mtok covers cost. A latency-niche player facing commoditized token pricing and unproven gross margins can bleed even with cash in the bank.

Pre-mortem (18 months out, thesis broke): GroqCloud revenue stayed sub-scale and unprofitable per-token; the Saudi build never re-ramped; Nvidia shipped licensed-LPU-inspired inference SKUs that ate the latency niche; the $650M raise funded a slow burn rather than a breakout; secondaries drifted toward the cash value of the stub. The company becomes a well-funded zombie or a quiet absorption.

Are multiples too high? There's no earnings multiple. The relevant valuation check is the secondary markdown to ~$2.8B from $6.9B — the market has already repriced ~60% lower. So the surprise risk is arguably symmetric here: the bad news (brain-drain, Saudi miss, licence) is substantially in the mark.

Contrarian view (what the market refuses to see): Consensus reads the Nvidia deal as a humiliation/end-of-story. The contrarian read is that management converted a deteriorating, capital-intensive, single-customer hardware business into ~$17B of cash at the top — and what's left is a real, cash-rich, niche-leading inference cloud trading at a steep discount. If GroqCloud proves any durable unit economics, the marked-down stub could be mispriced to the upside. That is a genuine 2-sided debate — which is exactly why direction here is WATCHING, not a conviction call.

Lens 13 · Devil's Advocate (short-seller)

Dismantling the bull case:

• What structurally breaks the money machine? Token-price commoditization (DeepSeek −75%) collides with a high-host-overhead, SRAM-fan-out architecture that's "not competitive for throughput". If buyers optimize $/token instead of latency, Groq's edge is irrelevant for the majority of inference spend.
• Where's revenue concentrated, and what if it shifts? It already shifted — Saudi → Nvidia/AMD, a $1.5B hole. The replacement is a low-monetization developer base.
• Why is the moat weaker than bulls think? Because Groq itself sold a perpetual licence to it and its three-lines-of-code switching cost is ~zero in both directions.
• Most dangerous competitor bulls underestimate: Nvidia — now armed with Groq's own licensed IP and its team, with the distribution, capital and software (CUDA/TensorRT-LLM) to commoditize the latency niche from above. Secondarily, hyperscaler in-house silicon (TPU v6, Trainium) that bundles inference into the cloud the developers already use.
• Worst capital-allocation / governance read: the deal can be framed as management+early-investors cashing out $7.6B and shipping the talent to Nvidia while leaving new secondary buyers holding a hollowed stub — and the DOJ/FTC antitrust inquiry means even the cash thesis carries unwind risk.
• What must hold for today's mark? That GroqCloud reaches standalone positive unit economics and re-wins large recurring demand before the cash cushion erodes — neither yet evidenced.
• −20–30% growth scenario: with no earnings to compress, a growth disappointment simply walks the secondary mark toward the net cash value of the stub — plausibly below today's ~$2.8B context if burn accelerates.
• Single permanent-impairment scenario, and plausibility: Nvidia ships licensed-tech inference products that erase the latency differentiator while token prices commoditize the rest — GroqCloud never clears profitability and winds down to its cash. Plausibility: medium-high.

Lens 14 · Management Questions (ordered by information value)

1. What is GroqCloud's recurring revenue run-rate and gross margin, fully excluding the Nvidia licensing payments? (The whole thesis turns on this one number.)
2. At $0.27/Mtok, are you gross-margin-positive per token today on current-gen 14nm hardware — and what does that become on Samsung 4nm?
3. Post-deal, how many of your ~90%-departed engineers' roles have you back-filled, and how is the next-gen LPU roadmap staffed now that Ross is at Nvidia?
4. What replaces the ~$1.5B Saudi hardware revenue that slipped — name the recurring customers and their committed volumes.
5. Nvidia holds a perpetual licence to your architecture. Concretely, what stops Nvidia from shipping a competing low-latency inference product that commoditizes your niche?
6. How exposed is the ~$17B cash consideration to the DOJ/FTC antitrust review of the deal — are any installments contingent or clawback-able?
7. Why distribute $7.6B to shareholders rather than retain it to fund GroqCloud's scale-up — what does that signal about the cloud opportunity you see?
8. With DeepSeek and others cutting token prices ~75%, what is your pricing/cost roadmap to stay viable as inference commoditizes?
9. What is your current monthly burn, and how many months of runway do the $650M + $500M raises actually buy?
10. What share of GroqCloud's ~2M developers are paying, and what is net revenue retention on paying accounts?
11. Is the Dammam/Aramco deployment generating revenue today, and what are the contracted economics now that KSA also buys Nvidia/AMD?
12. What is the realistic path and timeline to an S-1, given multiple analysts now read an IPO as less likely post-deal?
13. How dependent is your next-gen roadmap on Samsung SF4X yield, and what's the contingency if 4nm volume slips again?
14. What is the durable use-case where latency/determinism is worth a price premium large enough to offset your higher chip-count cost — and how big is that TAM?
15. Eighteen months out, what is the single metric on which you'd want this management team judged?

---