CBRS IC Memo | IWANNAVY LAB

Executive Summary & Action Plan

Verdict

SELL — Cerebras materially raises OpenAI's inference ceiling and makes real-time Codex practical; OpenAI's own networking and API work transfers that speed to users, while any improvement in model intelligence or training remains unproven. The >$20B relationship is commercially transformative, but $215.08 discounts an exceptional 750MW ramp before margin, utilization, delivery, and dilution evidence exists.

The stock sits above all three intrinsic-value methods. The updated $86 fair value uses a 301M committed-capacity dilution case; the $280 market Bull endpoint is a premium-persistence scenario, not fair value.

Loading chart

Price $215.08 | Market cap $47.931B | Target $86.00 | Upside -60.01% IWANNAVY Fair Value · price as of 2026-07-10 regular close / 2026-07-11 05:00 KST · two market feeds $215.08 / $215.08 · provider basic market cap; 301M committed-case diluted value ≈$64.7B · Street $291.09 (n=11, reference only)

The operating dashboard separates what OpenAI has already validated from what the stock still assumes.

Loading chart

Abstract

Cerebras already makes one OpenAI product materially faster, but it has not been shown to make OpenAI models smarter. GPT‑5.3‑Codex‑Spark runs on WSE‑3 above 1,000 tokens per second, with production bursts to 4,000; Spark is a smaller model and OpenAI separately reduced API and network latency, so no same-weight test isolates the hardware effect. Commercially, OpenAI committed to 750MW worth more than $20B, yet Q1 MRA revenue was only $16.9M. The $25B RPO is not equivalent to high-margin backlog: initial-250MW pass-through costs are included, customer warrants reduce revenue, and financing interest increases deferred revenue. We separate $86 intrinsic value from a $144 12M market Base. A 39-close block bootstrap supplies volatility bands, three age-matched IPOs test premium persistence, and delivery/margin scenarios set prices. The resulting 12M expected price is $136.75, 36% below the $215.08 close. SELL above $185; revisit after 250MW and margin proof or below $108.

Forecast Path

The expected path is almost flat for one month, bottoms around the supply/250MW decision window, then recovers modestly as delivery evidence becomes observable. The 12M expected value is $136.75 (-36.4%); the shaded band is a 70% volatility interval, not a scenario range.

Loading chart

Horizon	Bear / Base / Bull	Judgmental probabilities	Expected	70% volatility band	Expected return
1M	$159 / $214 / $284	30% / 50% / 20%	$211.50	$156.90–284.28	-1.7%
3M	$125 / $178 / $280	35% / 50% / 15%	$174.75	$104.77–291.73	-18.7%
6M	$75 / $135 / $260	35% / 50% / 15%	$132.75	$64.38–273.22	-38.3%
12M	$65 / $144 / $280	35% / 50% / 15%	$136.75	$49.46–384.03	-36.4%

Probabilities are an uncalibrated judgmental overlay on the empirical distributions below; their roundness is deliberate and avoids false precision.

Action Plan

Do not chase the post-announcement rally; trim or exit above the $185 IPO price and re-enter only in stages after evidence or at a valuation discount.

Loading chart

12M Bull/Bear reward-to-risk 0.43:1 (+30.2% to $280 vs -69.8% to $65) | Prob-weighted 12M return -36.4% | P(thesis wrong) 32%, judgmental | Confidence: Medium (model-risk overlay: low)

Forecast Drivers & Expectations

Observed Cerebras contribution — inference throughput: OpenAI says specialized Cerebras hardware enabled Codex‑Spark's >1,000 tokens/s target; production traffic reached 1,000 and burst to 4,000. WSE‑3 provides a latency-first decode tier inside OpenAI's production stack (OpenAI engineering, 2026-04-22).
Shared end-to-end result — not a hardware A/B: the roughly 65→1,000 tokens/s comparison spans different models; Spark is a smaller GPT‑5.3‑Codex variant. Total task time also includes prefill, tools, and network, so Cerebras' exact causal share is not publicly separable (Codex‑Spark launch, 2026-02-12).
Observed OpenAI contribution — transport and harness: persistent WebSocket plus Responses API work cut client/server roundtrip overhead 80%, per-token overhead 30%, and time-to-first-token 50%; a later production review reported agent loops 40% faster end-to-end. Do not credit these gains to Cerebras.
Unproven contribution — intelligence and training: OpenAI published strong benchmark performance but no same-weight GPU-versus-Cerebras quality test, no accuracy lift caused by hardware, and no evidence that current OpenAI weights were trained on Cerebras. The signed 750MW commitment is explicitly inference capacity; frontier-model/hardware co-design is future option value, not a realized quality KPI.
Commercial causal path: Q1 OpenAI MRA revenue was $16.9M; committed capacity steps to 250MW by end-2026, 500MW by end-2027, and 750MW by end-2028. The additional 1.25GW is an option, not backlog certainty (MRA Exhibit 10.11).
RPO-quality bridge: $25.0B is gross RPO, not gross profit. Initial-250MW rent, power, and related pass-through costs are included and recognized gross. Q1 pass-through revenue/cost were $4.111M/$3.991M—only $0.120M gross profit, or 2.9%—while the $516.1M customer-warrant asset is amortized as a reduction of revenue through 2031 (Q1 10-Q).
Financing bridge: Q1's $18.9M non-cash financing-component interest increased deferred revenue; future GAAP revenue therefore does not map one-for-one to contemporaneous cash receipts. This and pass-through mix are why MW and RPO cannot be capitalized at a single high-margin multiple.
Dilution bridge: 226.533M basic shares plus treasury-stock options, RSUs, AWS warrants, and an estimated committed-750MW OpenAI warrant case produce about 301.0M shares; the 2GW maximum reaches about 315.6M before 9M performance RSUs. Current diluted equity value is therefore roughly $64.7–67.9B, not $59.2B.
Expectations gap: $215.08 implies roughly 15.7x FY28 Base EV/Sales on the 301M share case, before Cerebras has reported 250MW availability or sustained normalized margin.
Outside views: a centered three-day block bootstrap uses 38 post-IPO returns only for dispersion; an age-matched CRDO/ALAB/CRWV panel uses three observations only as a premium-persistence check. Neither is large enough to claim calibrated probabilities.

Forecast Evidence Stack

Evidence layer	Reproducible input	What it says	How it is used
CBRS time series	39 closes / 38 returns, 2026-05-14–07-10	112.2% winsorized annualized volatility; -45.8% max drawdown	70% bands only; sample mean removed
Age-matched AI IPOs	CRDO, ALAB, CRWV from each 39th close; n=3	Median 1M/3M/6M/12M returns: -11.6% / -7.9% / -13.2% / +5.3%	Rejects forced 12M convergence to IFV
Intrinsic value	DCF, FY28 EV/EBITDA, FY27 EV/Sales	$86 median on 301M shares; $82 at 315.6M	Long-run anchor, not 12M Base
Contract economics	RPO timing, pass-through, warrants, financing component	Scale is contracted; margin and cash conversion are not	Discounts high-multiple revenue
Event path	Q2 results, 138.4M eligibility, 250MW gate	6M is the maximum-information-risk point	Drives the U-shaped expected path

The bootstrap winsorizes daily log returns at the 5th/95th percentiles, removes the sample mean, resamples circular three-day blocks across 200,000 paths with seed 20260711, and applies q15/q85 multipliers to each horizon's scenario-weighted target. The analogue panel is explicitly not called a base rate.

Disclosed Period Coverage

Only two discrete quarters are public; the prospectus supplies Q2+Q3'25 in combination and Q4 revenue only. Revenue nearly doubled YoY and operating loss narrowed, but Q1'26 free cash flow remained -$119.6M after $132.0M capex (Q1 10-Q).

Period	Revenue ($M)	Gross margin	Operating margin	Net income ($M)	FCF ($M)	Disclosure quality
Q1'25	99.5	41.8%	-28.8%	-23.9	-153.2	Full comparative quarter
Q2+Q3'25	239.1 combined	—	—	—	—	Two-quarter aggregate only
Q4'25	171.4	—	—	—	—	Revenue-only bridge
Q1'26	193.4	44.6%	-8.4%	-14.0	-119.6	Full 10-Q quarter

Comparable Q1 periods show improving accounting margins but still extreme investment cash burn.

Loading chart

Revenue & Profit Mix

Hardware remains larger, while cloud and services is the faster-growing line and the channel through which OpenAI capacity monetizes.

Loading chart

Q1'26 was near operating breakeven but still loss-making on GAAP; the next debate is whether datacenter utilization can offset the guided Q2 gross-margin step-down.

Loading chart

Business Lines

Wafer-scale systems (57.2% of Q1 revenue): CS‑3/WSE‑3 combines roughly 900k cores, 44GB on-chip memory, and 21PB/s bandwidth. That architecture removes inter-GPU memory traffic from decode, but system shipments remain lumpy and depend on manufacturing and datacenter readiness.
Cloud & other services (42.8%): $82.8M revenue grew 178% YoY. This is OpenAI's economic lane and the source of recurring revenue, but utilization starts below full load while lease, power, and rent-back costs arrive early.
Strategic capacity: OpenAI committed 750MW and first public production is Codex‑Spark; Europe plans first capacity by end-2026 and 200MW by end-2027, partly for OpenAI. Flex plans ~7x CS‑3 manufacturing capacity, but capacity is not shipment, revenue, or profit (Europe; Flex).
Relationship economics: OpenAI is customer, secured lender, and potential major shareholder: a ~$1.0B working-capital loan, 6% rate, and up to 33.445M near-zero-strike Class N warrants align delivery but concentrate revenue, liquidity, and dilution risk.

IWANNAVY Fair Value

The prior operating model's enterprise-value numerators are unchanged, but the per-share denominator is corrected. Replacing the legacy 277M basis with a 301M committed-750MW dilution case moves DCF / EV-EBITDA / EV-Sales from roughly $93 / $103 / $91 to $86 / $95 / $83. The robust median is $86; the 315.6M maximum-case sensitivity is about $82. EV/FCF remains unusable while forward FCF is negative.

All three methods remain far below spot; Street's $291 mean is a reference, not an input, and embeds faster backlog conversion plus a persistent scarcity multiple.

Loading chart

Intrinsic range: Bear $53 / Base $86 / Bull $108. The 12M market path is separate: Bear $65 / Base $144 / Bull $280.
Critical correction: Q1's 63% customer was not OpenAI; the 10-Q identifies MBZUAI at 63% and G42 at 11%, while the OpenAI MRA produced $16.9M. Present revenue is UAE-concentrated; future RPO is OpenAI-concentrated.

12M Market-Price Valuation Bridge

The market-price scenarios use (FY28 revenue × EV/Sales + FY27E net cash) / diluted shares; they are forward market states, not three additional intrinsic-value methods.

Scenario	Operating condition	FY27 revenue	FY28 revenue	FY28 EV/Sales	Net cash	Diluted shares	12M price
Bear	Q2 revenue <$190M or GM <35%; 250MW delay	$1.35B	$2.4B	6x	$4.5B	290.9M	$65
Base	Q2 $190–200M; GM 36–38%; on-time 250MW	$1.9B	$3.8B	10x	$5.2B	301.0M	$144
Bull	Q2 ≥$205M; GM ≥40%; utilization/RPO lead	$2.6B	$5.5B	15x	$6.0B	315.6M	$280

Dilution bridge	Incremental shares	Treatment
Basic Class A/B/N	226.533M	Reported at 2026-06-17
Options, treasury-stock method	26.590M	Modeled incremental shares
Regular RSUs	27.629M	Included
AWS warrant	1.435M	Maximum modeled increment
OpenAI warrant, committed-750MW case	18.813M	Vested/market conditions plus proportional capacity estimate
Committed-case total	301.0M	Headline IFV and Base scenario
Additional OpenAI 2GW-option sensitivity	+14.632M	315.6M max, before 9M performance RSUs

Catalysts, Leading Indicators & Risks

Window	Observable	Bull trigger	Bear trigger	Forecast action
Q3'26	Q2 core revenue / GM	≥$205M / ≥40%	<$190M / <35%	Move 10 points between Bear and Bull
Q2 result–10/28	Up to 138.4M shares become sale-eligible	Holds >$185 on normal turnover	Form 144/Form 4 plus high-volume break	Reweight 3M/6M only on observed sales
By 12/31	Initial OpenAI 250MW availability	On time; utilization disclosed	Delay, SLA issue, or reduced scope	Rebuild 6M/12M revenue path
Quarterly	OpenAI MRA revenue / core gross profit	Sequential step-up with clean GP	Pass-through rises without core GP	Cut revenue multiple before revenue
Quarterly	Gross RPO conversion	At/above 16% cadence with margin	Recognition shifts right or stays low-margin	Lower Base multiple and net cash
By 12/31	First European capacity	Energized site with live workload	Schedule/economics remain undisclosed	Option value only until live

Falsification & Review Rules

SELL-thesis falsifier: Q2 core GM at or above 38%, on-time initial 250MW availability by 2026-12-31, clear OpenAI revenue acceleration, and absorption of staged share supply above $245. If all occur, rebuild the DCF and Bull multiple; do not auto-buy on one datapoint.
Downside confirmation: Q2 core GM below 35%, a disclosed OpenAI committed-date delay, or RPO recognition pushed beyond the 16%/45% cadence. Exit residual speculative longs and raise Bear weight.
Price control: $185 is the no-chase line, not a thesis falsifier. For entries below $108, a sustained close below $53 is the severe-failure stop; operational evidence remains the thesis test.
Review cadence: weekly price/supply check; immediate review on earnings-date announcement, Q2 results, capacity-status disclosure, or an OpenAI product expansion; formal refresh by 2026-10-11.

References

OpenAI–Cerebras partnership (OpenAI, 2026-01-14) · GPT‑5.3‑Codex‑Spark (OpenAI, 2026-02-12) · WebSocket production review (OpenAI, 2026-04-22)
MRA Exhibit 10.11 (SEC, 2026-05-04) · 424B4 prospectus (SEC, 2026-05-14)
Q1'26 results (2026-06-23) · Q1'26 10-Q (filed 2026-06-24)
Europe 200MW plan · Flex 7x manufacturing plan (2026-07-09)
CBRS history (S&P Global market data) · Nasdaq CBRS · Street target distribution
Age-matched histories: CRDO · ALAB · CRWV; inclusion checks: CRDO S-1 · ALAB 424B4 · CRWV 424B4