Period ended 30 June 2026; filed August 2026 · Primary: financial statements; Notes 3, 10, 12–15 and 17; MD&A, liquidity, customer contracts and risk factors. Financial actuals stop at June; common shares on the cover are dated 5 August.
https://www.sec.gov/Archives/edgar/data/2021728/000162828026056357/cbrs-20260630.htmSpeculative watchlist · financing not fully underwritten
The investment decision
At $127.21, our demanding base illustrates 11.9% annualised over five model years. A ~$90 conditional 20% hurdle ceiling depends on a 25x terminal cash multiple; the central perpetual-growth DCF is only $59.31. Contract conversion, funding and lifecycle capex must earn confidence before the stock earns conviction.
Edition update. Initial coverage using Q2 2026 filings, primary product materials and four supplied analyst notes. Stale note prices and targets are excluded. Model and source distinctions are explicit; no prior report is overwritten.
01. Faster answers are useful; profitable capacity is the investment
Cerebras sells AI computing systems and operates inference capacity that customers use as a service. Its wafer-scale design keeps a very large amount of computation and high-speed memory on one silicon wafer. Shorter data journeys can reduce the delay between tokens when an AI model writes an answer. That is commercially valuable for interactive products and agents waiting for their next step. It is not proof that this architecture wins every workload, every model or every cost comparison. [S1] [S4] [S5]
The system can complement conventional accelerators: other chips can process the initial prompt while Cerebras handles rapid token generation. CS-4 is a system-level development involving multiple wafers, networking, power and cooling. Company speed comparisons remain configuration-specific claims; this report does not turn “up to” performance or extrapolated model sizes into proven customer economics. The competitive test is delivered cost per useful response at the required latency, reliability and security, not a single laboratory speed figure.
There are two different ways to monetise the technology. A hardware sale transfers a system to a customer and can produce earlier cash without Cerebras operating that capacity for years. Cloud service can produce recurring revenue and a larger lifetime opportunity, but Cerebras must procure equipment, secure power, construct capacity and pay operating costs before collecting all the revenue. Selling more cloud capacity can therefore increase both growth and near-term cash needs. Cloud share of Q2 GAAP revenue was 70%, versus 32% a year earlier. [S1]
Our position is a speculative watchlist, with medium evidence confidence and an incomplete financing underwriting case. The product and contracts support a real opportunity. The current price still requires a large, profitable build-out, while full future funding terms, mature utilisation and lifecycle replacement costs remain unproven. A share-price decline or a large contract number alone does not establish a bargain.
02. A $25.4bn contract book is a delivery obligation
The June filing reports $25.4bn of remaining performance obligations, or RPO: contracted revenue not yet recognised. About 22% is expected within the first 24 months, 43% in the following 24 months and 35% thereafter. Those proportions imply approximately $5.59bn, $10.92bn and $8.89bn respectively. RPO includes some pass-through amounts and a significant OpenAI commitment. It is neither present cash nor a gross-profit measure. [S1]
OpenAI has committed to 750MW of capacity delivered in tranches during 2026–28, with an option for a further 1.25GW through 2030. The optional amount must not be described as secured demand. Separately, management described more than 600MW of live and contracted capacity for delivery by the end of 2027. These numbers have different scopes and dates: “live and contracted” does not mean already energised. Failure to meet delivery or service requirements can allow contract termination. [S1] [S2]
The cash loop matters. Infrastructure must be available, customers must accept and use the contracted service, and billings must become cash after service credits. The OpenAI working-capital loan originally provided about $1.00bn; $918.2m remained at June. It can be repaid through service fees, compute, hardware and other permitted credits as well as cash. June loan offsets already included $86.3m of non-cash billings. Such financing helps build the business today while reducing cash available from some future service delivery. [S1]
Demand is concentrated. The three largest customers represented 34%, 32% and 10% of Q2 revenue, together 76%. Two customers represented 76% of receivables. Consequently, a small number of counterparties have substantial influence over delivery schedules, pricing, payment and technology deployment. The report treats customer security restrictions and the possibility of workloads remaining on a customer’s own infrastructure as risks; analyst channel checks about specific frontier-weight decisions are not confirmed cancellations. [S1] [S7]
03. Growth, adjustments and cash tell different stories
| USD million unless stated | Q2 2025 actual | Q2 2026 actual |
|---|---|---|
| Hardware GAAP revenue | 70.295 | 54.119 |
| Cloud and other GAAP revenue | 33.027 | 125.991 |
| Total GAAP revenue | 103.322 | 180.110 |
| GAAP gross profit | 32.099 | 25.559 |
| GAAP gross margin | 31.1% | 14.2% |
| GAAP operating profit / (loss) | (57.182) | (477.233) |
Q2 GAAP revenue rose 74%, driven by cloud growth of 281%; hardware revenue declined 23%. The gross-profit outcome did not keep pace. Customer warrants, pass-through costs and the IPO compensation event make a simple year-on-year margin or earnings comparison misleading. [S1] [S2]
| Q2 revenue reconciliation | USD million |
|---|---|
| GAAP revenue | 180.110 |
| Add customer-warrant amortisation | 44.262 |
| Remove pass-through revenue | (14.503) |
| Management core revenue | 209.869 |
Management’s core revenue removes specific accounting items; it is not additional cash received. Core gross margin was about 40.6%, versus 14.2% under GAAP, while core operating loss was $33.6m. Customer warrants are economically meaningful because customers can gain equity and existing holders can be diluted. A stronger adjusted margin is useful for understanding operations, but ignoring the equity claim would overstate the shareholder benefit. [S2]
Q2 stock compensation was $377.0m, including $273.6m associated with the IPO liquidity condition. This one-time vesting event should not simply be annualised. Future compensation still has a cost. Our cash model expresses that cost through an expanded share denominator and further annual dilution; it does not also subtract the same non-cash expense from cash. The prior-year net-profit comparison is also distorted by a $363.3m gain from extinguishing a forward-contract liability, so a historical P/E is not informative. [S1]
| H1 2026 cash bridge — actual | USD million |
|---|---|
| GAAP revenue | 373.516 |
| Cash used in operations | (47.488) |
| Cash purchases of property and equipment | (548.873) |
| Operating cash less cash capex | (596.361) |
The $596.4m cash gap is calculated from the same six-month period. It excludes purchases of investments and is not an adjusted-earnings figure. Customer deposits, deferred revenue and working-capital movements influence operating cash, while capex payables are not yet cash spending. The correct question is whether subsequent capacity produces enough cash to cover equipment replacement, rent, power and financing—not merely whether revenue keeps rising. [S1]
04. A substantial cash balance, with substantial claims on it
| 30 June 2026 balance / commitment | USD bn | Treatment |
|---|---|---|
| Cash plus investments | 7.922 | Unrestricted liquidity starting point |
| Restricted cash | 0.685 | Excluded from model available cash |
| Customer working-capital loan | 0.918 | Fully reserved before forecasting |
| Operating lease liability, present value | 0.569 | Rent remains in operating costs |
| Commenced future lease payments, undiscounted | 0.690 | Not added again to debt after expensing rent |
| Additional uncommenced leases, including subsequent agreements | 2.300 | Future rent must be absorbed by margins |
| Undrawn revolving facility | 0.850 | Availability is not cash |
The May IPO raised approximately $6.2bn net, making external capital the main source of today’s liquidity. The $850m revolving facility had no June borrowings and matures on 14 April 2031. The filed post-IPO pricing is SOFR plus 2.25%, potentially falling to 2.00% on the specified leverage test; our future incremental funding rates of 7.5–12% are house assumptions, not the facility’s current contractual rate. [S1]
Our opening model reserve is deliberately conservative: $7.922bn unrestricted liquidity less the whole $0.918bn customer loan and an assumed $2.0bn second-half cash use leaves $5.003bn for the five-year path. The $2.0bn is a Capital Thesis planning assumption, not a reported balance or company guidance. Restricted cash remains excluded even though some restricted resources can support loan-related obligations; this conservatism can understate available economic liquidity. We reserve the customer loan once and do not deduct its future service-credit repayment a second time.
Rent and power are included in forecast cash operating costs. Therefore the model does not also deduct operating-lease liabilities as financial debt. The $2.3bn uncommenced lease total is an important execution test for that margin assumption, not a second cash subtraction. The model has no annual lease-cohort or equipment-retirement schedule; a detailed project-level build could reveal greater cash needs.
Even this opening liquidity does not make the plan self-financing. The base path requires about $1.11bn of peak new debt, above the currently disclosed $850m facility. The bull reaches $1.76bn and the bear $3.02bn. This is a model funding plug, not verified lending capacity. If additional financing is unavailable, or lenders demand more equity, the forecast must be rebuilt with slower capacity growth, higher dilution or distress. That unresolved access to capital is a reason for the watchlist stance.
05. The share count cannot stop at the IPO
The latest filed common-share count on the cover is 237.564m as of 5 August. At the $127.21 reference price, that gives a dated basic market-capitalisation proxy of $30.22bn. It is not a fully diluted valuation or a live October share count. The quarterly weighted-average EPS denominator is unsuitable for this purpose. [S1] [S3]
| Consistent June economic share bridge | Million shares |
|---|---|
| June common classes A, B and N | 228.204 |
| All OpenAI warrant shares, including conditional tranches | 33.445 |
| All AWS warrant shares | 2.697 |
| Outstanding options | 23.797 |
| Unvested RSUs | 16.532 |
| Performance RSUs | 9.000 |
| Gross model opening economic claims | 313.675 |
The model assumes all these contingent warrants and awards ultimately become shares, ignores exercise proceeds and adds future new-award dilution of 3.5%, 2.0% or 1.5% annually in bear, base and bull. This is a conservative gross-claim convention, not the GAAP treasury-stock method or a forecast that every performance hurdle will be earned. It can overstate dilution in a weak outcome. The annual allowance covers awards beyond the pool already counted; it is not a second charge for the same grants.
The 10.034m OpenAI warrants exercised in July are already inside the full 33.445m warrant pool; they are not added again through the August common count. Earlier G42 warrant exercises are already in June common shares. Early-exercised shares subject to repurchase are also not added again. This consistent June bridge avoids mixing current common shares with warrants that have since changed form. There is no complete October award roll-forward in the reviewed filings. [S1]
Multiple share classes, customer-linked equity and large performance awards make per-share results more uncertain than revenue results. Governance and incentive alignment therefore belong in the underwriting. The company also reported material weaknesses in financial-reporting controls at June. Remediation progress is a monitoring item, not evidence of misconduct. [S1]
06. The five-year model pays for the equipment
Capital Thesis uses a cash-based enterprise valuation because an earnings multiple that ignores the capacity bill is poorly suited to this build-out. FY2027–31 represent five stylised annual periods from the dated purchase price; they are not an exact calendar-date IRR. All following forecast figures are our independent assumptions, not management guidance or broker targets. Forecast sales use an economic, core-like basis: pass-through revenue is excluded and non-cash customer-warrant contra-revenue is reversed. The customer equity cost is captured through dilution; this is not a formal GAAP revenue forecast.
| Base house path | FY27 | FY28 | FY29 | FY30 | FY31 |
|---|---|---|---|---|---|
| Revenue, $bn | 3.00 | 7.00 | 11.00 | 15.00 | 20.00 |
| Cash operating margin | 15% | 32% | 40% | 42% | 44% |
| Cash capex, $bn | 3.00 | 4.20 | 3.85 | 4.05 | 4.40 |
| After-capex unlevered cash, $bn | -2.613 | -2.265 | -0.102 | 1.272 | 2.930 |
| Closing new debt, $bn | 0.000 | 0.916 | 1.110 | 0.000 | 0.000 |
| Economic shares, m | 319.95 | 326.35 | 332.87 | 339.53 | 346.32 |
The mechanism is explicit: revenue times cash operating margin, less cash taxes, capex and incremental working capital, equals after-capex unlevered cash. Cash operating margin is before depreciation, non-cash stock compensation, taxes and financing interest, but after cash operating expenses including rent and power. Taxes are 22% of positive cash operating profit less assumed depreciation. No NOL benefit, interest tax shield or interest income is credited. Working capital absorbs 3% of incremental revenue.
The base requires revenue to rise from a FY26 core-guidance midpoint near $0.885bn to $20bn in FY31, with cash operating margin reaching 44%. Annual capex falls from 100% of revenue in FY27 to 22% in FY31; cumulative five-year capex is $19.50bn. Depreciation is assumed to fall from 20% to 14% of revenue. This is a demanding scale-and-efficiency story: contractual demand alone cannot deliver the assumed utilisation, pricing and replacement economics.
A $1bn minimum cash balance is maintained. Deficits are funded with new debt, interest is charged on average annual debt, and excess cash first repays that debt. Shares compound using the case-specific new-dilution allowance. No dividends, buybacks, acquisitions or new external equity issuance are assumed. A project-specific financing or equity raise would change these results; the current model is not a complete three-statement forecast.
For comparison, management’s August FY26 core revenue guidance was $880–890m, core gross margin 41–43% and core operating margin negative 19–17%; it expected revenue to more than triple in 2027. A supplied August analyst model projected $3.172bn for 2027 and $9.271bn for 2028, while another supplied framework used about $6.3bn for 2028. Our $3bn/$7bn base is independent and does not inherit either target. These sources illustrate forecast disagreement rather than an established consensus. [S2] [S7] [S8] [S10]
07. Conditional outcomes, not target-price probabilities
| House case | FY31 revenue | Cash op margin | Capex / sales | Exit cash multiple | Terminal / share | 5Y annual return |
|---|---|---|---|---|---|---|
| Bear | $8bn | 30% | 25% | 18x | $0.00 | Total loss |
| Base | $20bn | 44% | 22% | 25x | $223.18 | 11.9% |
| Bull | $35bn | 50% | 20% | 30x | $692.15 | 40.3% |
The bear still grows revenue to $8bn but converts too little into cash: a 30% cash operating margin and 25% capex ratio leave only $93.6m terminal unlevered cash after taxes and working capital. At 18 times that cash, enterprise value is below the debt claim after available cash. Common equity is floored at zero. This is a conditional total-loss illustration, not a default probability or the only possible adverse path.
The base FY31 bridge is $20bn revenue × 44% = $8.80bn cash operating profit, less $4.40bn capex, $1.32bn tax and $0.15bn working capital = $2.93bn after-capex unlevered cash. At 25 times, enterprise value is $73.25bn. Add $4.042bn cash and subtract zero remaining new debt, then divide by 346.322m economic shares: $223.18 per share. From $127.21, the five-year annualised return is 11.9%.
The bull reaches $35bn revenue, a 50% cash operating margin and a 20% capex ratio. At 30 times $7.428bn terminal cash, plus accumulated balance-sheet cash and with 337.917m shares, terminal value is $692.15 and the annualised return 40.3%. It requires faster deployment and higher utilisation, strong pricing and operating leverage, lower future dilution and a premium exit valuation simultaneously. It is an upside boundary, not a central expectation.
Five-year annual return = (terminal share value / $127.21)^(1/5) − 1; no distributions are included. Returns are nominal USD before investor tax, fees and currency effects. No scenario probabilities are assigned. A positive current free-cash-flow multiple is unavailable, so a meaningful like-for-like “no rerating” return cannot be calculated. This report intentionally remains outside the website’s comparable-return ranking rather than inventing that input.
08. The valuation cross-check argues for restraint
The 25-times cash multiple is a major assumption. It corresponds to a 4% terminal unlevered cash yield and still requires durable growth beyond FY31. A simplified perpetual-growth DCF of the same base cash path gives only $59.31 per share at a 13% discount rate and 3.5% perpetual growth. Across 11–15% discount rates and 2.5–4.5% growth, the range is $43.54–90.28. These are present-value cross-checks, unlike the future scenario prices.
The DCF adds the opening $5.003bn reserved net cash once and uses the base terminal diluted share count throughout, conservatively charging future dilution up front. It does not add future balance-sheet cash a second time. The perpetuity approach assumes cash growth slows sharply after FY31; the 25x exit approach rewards a longer growth duration. Neither is a verified fair value, and their disagreement is material rather than an error to average away.
| One change to base; all others held fixed | Terminal / share | Annual return | 20% hurdle ceiling |
|---|---|---|---|
| Base | $223.18 | 11.9% | $89.69 |
| Capex +5 percentage points of sales, every year | $141.96 | 2.2% | $57.05 |
| Cash operating margin −5 percentage points, every year | $159.72 | 4.7% | $64.19 |
| Annual new dilution +1 percentage point | $212.55 | 10.8% | $85.42 |
| Incremental debt rate +3 percentage points | $222.98 | 11.9% | $89.61 |
| 20x terminal cash rather than 25x | $180.88 | 7.3% | $72.69 |
Equipment intensity matters more than a modest change in borrowing rates in this particular path. Raising capex by five percentage points of annual revenue cuts the base return from 11.9% to 2.2%, even after recalculating funding and interest. That is why project returns, usable capacity and replacement spending matter more than the headline revenue multiple. A 20x exit cash multiple cuts the return to 7.3%.
The fixed-model price that would yield 20% annually is $223.18 / 1.20^5 = $89.69, approximately $90. It is a conditional hurdle ceiling, not an automatic buying trigger or independently established fair value. It holds the operating path, 25x exit cash multiple, financing availability and dilution schedule fixed. A lower trading price can change financing terms, award economics and demand expectations, none of which this reverse calculation reruns. The much lower central DCF cross-check reinforces the need to keep this distinction visible.
09. What would earn conviction—and what would break it
| Question | Evidence that would support the thesis | Evidence that would weaken it |
|---|---|---|
| Capacity converts to cash | On-time accepted deployment, utilisation and service collections consistent with the contracted schedule | Delays, acceptance disputes, credits or termination rights consume expected receipts |
| Product advantage survives competition | Repeat paid adoption at attractive all-in cost and latency across disclosed workloads | Conventional GPUs or custom chips close the useful cost/latency gap or security constraints shift mix adversely |
| Margins cover infrastructure | Cash costs, rent and replacement capex remain within the model while scale increases | Five-point capex overshoot or margin shortfall persists; current sensitivities show the value impact |
| Funding becomes underwritable | Committed capital sufficient for the project schedule with disclosed covenants and per-share consequences | The required facility expansion is unavailable, or equity issuance materially exceeds the model |
| Revenue broadens | More unaffiliated customers and a declining share of revenue/receivables from the largest counterparties | Large customers dominate pricing, service credits, cancellation exposure and equity claims |
| Reporting becomes clearer | Consistent GAAP/core bridges, award roll-forwards and control-remediation progress | Adjusted metrics obscure equity transfers, cash usage or changing performance obligations |
The next quarterly report and subsequent capacity-delivery updates are practical catalysts; no unverified earnings date is asserted. CS-4 deployments, manufacturing expansion and new sites can reduce delivery risk, but only customer acceptance and cash economics validate investment returns. Export restrictions, power availability, construction and supply-chain execution can delay monetisation even if demand is present. Product obsolescence can also make the assumed depreciation and capex path too optimistic. [S1] [S2] [S4] [S6]
The supplied four-note pack was used as a debate map, not an instruction set. The October Ultrafast note raises competitive and security questions; the September weekly contributes only dated CBRS forecast rows; the August Supernova note informs the architecture discussion after primary checks; and the August results note supplies a clearly labelled alternative forecast. Their USD177.65–262.06 prices, analyst targets, unrelated sector material and dated positioning statistics are not current inputs. No current short-interest, ownership-flow or options-crowding claim is made. [S7] [S8] [S9] [S10]
Research cutoff is 18 October 2026; the price is the 16 October regular close and financial actuals end 30 June. Evidence confidence is medium: filed historical figures and contract terms are strong evidence, while product comparisons, build-out economics and distant forecasts remain conditional. We have no project-level maintenance-capex study, complete October award roll-forward or committed funding plan covering the model’s peak debt. The appropriate stance is to watch the conversion from fast inference to cash available per share, then update assumptions—not to let a dramatic price chart substitute for that work.
Financial exhibits
Exact observations accompany every chart. Actuals, management guidance and Capital Thesis assumptions are labelled separately; the chart shapes do not imply probabilities.
Cloud growth changes the business and its capital needs
Q2 cloud and other revenue grew rapidly while hardware revenue declined; the two routes have different cash timing.
| Observation | Q2 2025 actual | Q2 2026 actual |
|---|---|---|
| Hardware | 70.295 | 54.119 |
| Cloud and other | 33.027 | 125.991 |
| Total GAAP revenue | 103.322 | 180.11 |
All figures GAAP and for the same three-month periods. Core revenue is separately reconciled in section 03. Sources: [S1] [S2]
The first-half capacity build consumed $596m after capex
Cash from operations was already negative before equipment purchases. Revenue growth is not current free cash flow.
| Observation | H1 2026 actual / arithmetic |
|---|---|
| Cash from operations | -47.488 |
| Cash capex outflow | -548.873 |
| Cash after capex | -596.361 |
Cash after capex is operating cash less cash purchases of property and equipment; not GAAP earnings or core profit. Sources: [S1]
The base only becomes cash-positive after the early build
Capex absorbs the operating gains until deployment scales; rent and power remain in cash operating costs.
| Observation | Cash operating profit | Cash capex | After-capex unlevered cash |
|---|---|---|---|
| 2027 house | 0.45 | 3 | -2.6134 |
| 2028 house | 2.24 | 4.2 | -2.2648 |
| 2029 house | 4.4 | 3.85 | -0.1024 |
| 2030 house | 6.3 | 4.05 | 1.272 |
| 2031 house | 8.8 | 4.4 | 2.93 |
Capital Thesis assumptions, not filed forecasts. Unlevered cash also deducts model taxes and incremental working capital; excludes financing interest. Underlying starting evidence: S1/S2. Sources: [S1] [S2]
All three cases assume substantial growth
Even the bear reaches $8bn sales: weak cash conversion and financing can still leave no equity value.
| Observation | Bear house revenue | Base house revenue | Bull house revenue |
|---|---|---|---|
| FY2027 | 1.8 | 3 | 4 |
| FY2028 | 3 | 7 | 10 |
| FY2029 | 4.5 | 11 | 17 |
| FY2030 | 6 | 15 | 25 |
| FY2031 | 8 | 20 | 35 |
Independent scenario assumptions; not probabilities or management guidance. Contracted revenue alone does not guarantee the model outcomes. Sources: [S1] [S2]
Source register and research basis
Evidence checked for this edition on 18 October 2026. Dates below identify source vintages. Company guidance and house scenarios are conditional; source links and the downloadable model allow the assumptions to be inspected.
12 August 2026 · Primary: quarterly GAAP/core reconciliations, Q3/FY26 guidance and management capacity and growth statements. Guidance and commercial claims are not audited outcomes.
https://www.sec.gov/Archives/edgar/data/2021728/000162828026056186/cbrsannouncesfinancialresu.htm16 October 2026 regular close; checked 18 October · Public market-data reference: USD127.21 at 4pm EDT. After-hours quotations are excluded. This is the last regular close before the research date.
https://stockanalysis.com/stocks/cbrs/history/18 August 2026 · Primary company product claims. Architecture and workload-specific performance require independent customer validation. Extrapolated maximum-model claims are excluded from the thesis.
https://www.cerebras.ai/blog/introducing-cerebras-cs-4Accessed 18 October 2026 · Primary product description. Performance comparisons depend on model, configuration and workload; they are not universal GPU comparisons.
https://www.cerebras.ai/cs4Accessed 18 October 2026 · Primary news and reporting index. The September Finland capacity announcement is a project milestone, not evidence that all capacity is live or earns target margins.
https://investors.cerebras.ai/1 October 2026; supplied research note, pp.1–3 · Secondary analyst commentary and channel checks about competing inference offerings, delivery and frontier-weight security. The linked URL identifies the publisher; the supplied note is not publicly hosted at that link. Its USD177.65 price and USD279 target are dated, not the present price or a Capital Thesis valuation.
https://www.morganstanley.com/28 September 2026; supplied research note, CBRS comparison rows · Mostly Micron/sector material. Only dated CBRS forecast/valuation rows inform the disagreement map; unrelated memory commentary and dated short-interest data are excluded. Publisher link is not a public copy of the note.
https://www.morganstanley.com/19 August 2026; supplied research note, pp.1–2 · Secondary technology interpretation, checked against company product materials. Stale USD220.01 share price and maximum-model extrapolations are not carried into the model. Publisher link is not a public copy of the note.
https://www.morganstanley.com/13 August 2026; supplied research note, pp.1–4 · Secondary 2026–28 forecasts, core-results reconciliation and target framework. Forecasts are labelled separately from filed actuals and house scenarios. Publisher link is not a public copy of the note.
https://www.home.barclays/