Blog · 2026-10-29 · Ashutosh Agarwal

An Agent Swarm on CoreWeave

A Matterfact agent swarm researched CoreWeave for two weeks. It withdrew its own first-day long, committed a fair value of $44.21 with a small underweight, then paused. Here is what it found, and what it got right and wrong once the market tested it.

CoreWeave (NASDAQ: CRWV). GPU cloud for AI, financed with about $41B of debt.

2 weeks of research, 17 Sep to 1 Oct; $44.21 vs $85.69 fair value vs share price, 29 Sep; −38% share price, 30 Sep to 23 Oct.

CoreWeave's value turns on one question: whether about $600B of capital spending over the next decade, financed largely with debt, will earn more than it costs. Because the equity sits on about $41B of debt, the answer moves its value between about $5 and $270 a share.

A Matterfact agent swarm that researched the company for two weeks concluded that new capacity will most likely earn little more than its cost of capital, which leaves little value for shareholders. On 29 September it committed a fair value of $44.21 a share, against a share price of $85.69, and took a small underweight with its loss capped in advance.

Three weeks later the stock was at $54.23, down 38% from its 30 September close, as credit markets repriced the company's debt. The swarm was right on direction and wrong on timing. It had paused for review on 1 October, it had expected the market to take far longer to get there, and several of its dated predictions are now running against it.

The process mattered as much as the call. The swarm withdrew its own first-day long within a day, caught two large errors in its own work, fixed in advance how it would be judged, and stopped when new evidence ran out. This article sets out what it concluded, the evidence behind the view, and what it got right and wrong once the market tested it. The swarm's full investment memo is below.

kind: word
title: CoreWeave investment memo
meta: Written by the agent swarm · as of 7 Oct 2026 · Word, 24 KB
href: /samples/coreweave-ic-memo-2026-10-07.docx

TL;DR

• The question was never growth. The swarm's revenue forecast matched the Street's almost exactly. What decides the value is whether about $600B of capital spending over ten years earns more than it costs to finance.
• Its first read was a long, and it withdrew it within a day. Fair value went from $98.74 to a committed $44.21, with a small underweight and a capped loss.
• It caught its own mistakes. A rule that worked in one direction only cut $57 a share in 6.5 hours with no new disclosure; the swarm found it and rewrote the rule. A review of its rebuild then removed $34 of margin counted twice.
• It fixed how it would be judged before the results. Tightening its own scoring rules cut the odds of adding to the underweight from 88% to about 17%, against its own bias.
• Right on direction, wrong on timing. The stock fell 38% in October as credit markets repriced the debt, a channel the swarm named at the start. It had expected the move to take much longer.

heading: What we cover
Why CoreWeave is hard to value :: one-probability-decides-coreweaves-value-because-its-equity-sits-on-about-41b-of-debt
The long that became a flat :: the-first-read-was-a-long-and-the-swarm-withdrew-it-within-a-day
The $57 mistake :: a-one-sided-rule-cut-57-a-share-in-6-5-hours-and-the-swarm-caught-it
The audit :: the-audit-confirmed-the-direction-rejected-the-level-and-changed-the-rules
The $34 double count :: a-review-of-the-rebuild-took-out-34-of-margin-counted-twice
What it found :: the-decisive-signals-were-in-contract-terms-credit-and-customer-concentration
Where it disagrees with the market :: the-question-was-never-growth-but-the-return-on-about-600b-of-borrowed-capital-spending
Why the stance was small :: a-low-fair-value-justified-only-a-small-underweight
How it will be judged, and the pause :: it-fixed-how-it-would-be-judged-before-the-results-then-stopped-when-the-evidence-ran-out
What happened in October :: in-october-the-market-repriced-coreweaves-financing
Right on direction, wrong on timing :: the-swarm-was-right-on-direction-and-wrong-on-timing
Questions for management and experts :: each-question-is-tied-to-a-test-set-in-advance
What this means for investment teams
The investment memo :: read-the-swarms-investment-memo

One probability decides CoreWeave's value, because its equity sits on about $41B of debt

CoreWeave's equity is worth about $270 a share if new capacity keeps earning well above its cost of capital, and about $5 if it earns only about its cost of capital. In the second case, growth adds costs as fast as it adds revenue and the debt absorbs the contracted cash flows. There is very little value in between.

CoreWeave rents out GPU computing capacity to AI labs and large technology companies. It buys the chips, leases powered data-centre buildings, signs multi-year contracts, and pays for nearly all of it with debt. That leaves shareholders with a thin slice of value on top of about $41B of debt.

Small changes in three things can move the equity from worthless to several hundred dollars a share:

1. What a GPU earns after its first contract ends.
2. What the next bond costs.
3. Whether a large customer renews.

So the whole valuation comes down to one probability: that new capacity keeps earning well above its cost of capital. A single discounted cash flow model, with one growth path and one discount rate defended to the end, does not fit a problem shaped like this.

What is an agent swarm?. A team of autonomous AI agents that work toward one research goal the way an investment team would: building the valuation, reading the filings, working through the credit and testing the thesis. It runs continuously and keeps score of its own predictions.

To value CoreWeave, the swarm assembled the evidence a sector specialist would want:

columns: 2
The backlog, customer by customer :: The contracted backlog by customer, and how much of it runs beyond four years. :: list
Contract pricing and length :: Disclosed pricing per megawatt and contract terms, checked against list prices per GPU-hour. :: target
The debt stack :: Bonds, convertible bonds (debt that can turn into shares), contract-backed loans, lenders' covenant tests, and liquidity against what falls due. :: layers
Credit markets :: Bond yields and default-swap prices (the cost of insuring its debt), read as a check on the equity model. :: network
GPU life and resale :: How long GPUs keep earning, and what older GPUs fetch per hour after their first contract. :: clock
Data centres and power :: Leased buildings, power capacity, and how much cooling and power equipment carries over to the next GPU generation. :: lens
NVIDIA ties :: NVIDIA's role behind customers, its financing of rival capacity, and the terms of its capacity backstop. :: compass
Insider and equity supply :: Executive share sales, the equity issuance programme and convertible bond issuance. :: users

The first read was a long, and the swarm withdrew it within a day

The swarm's first read was a long at $98.74 a share. Within a day it concluded that the upside rested on the bull case alone, and it went flat. The chart shows how fair value then swung in both directions as evidence arrived, while the share price held in the mid-$80s.

Fair value fell below the share price within a day of the first read, then swung widely CoreWeave, $ per share: the swarm's fair value at the end of each day it changed, against the daily closing share price, 16 to 29 September. Source: Matterfact agent swarm; daily closing prices.

CRWV share price: $83.35 on 16 Sep 2026, $85.92 on 29 Sep 2026 (10 observations).

17 September: $98.74. The first read was a conditional long: a small starter, about 2.6% of the portfolio, with a hedge against a sharp fall.

The same day: $82.41. The swarm's skeptic argued that financing is not a tail risk but a constraint in the base case. The swarm agreed. It raised the discount rate and modelled dilution explicitly, from 580M to 645M shares, and the apparent upside fell from about 23% to 2.5%.

By 18 September: the long is withdrawn. The remaining upside came only from the bull case: the median simulated outcome sat about 22% below the share price. The swarm went flat. It then added a rule that any future trade must be hedged against AI-chip stocks, which CoreWeave tracked with a sensitivity of 1.68 (a 1% move in those stocks has moved CoreWeave about 1.68%).

From then on, fair value moved in both directions as evidence arrived:

• It rose $10.83 when evidence on how long GPUs keep earning in their early years came in stronger than the bear case assumed.
• It rose $23.79 when the swarm found that its model had charged about $10.5B of first-half 2026 cash burn twice. A review confirmed the error.
• It fell when the loss of an anchor customer at renewal was written into the base case, and when the bull-case margin was cut toward that of listed data-centre landlords.

A one-sided rule cut $57 a share in 6.5 hours, and the swarm caught it

The largest move in fair value came from a rule that worked in one direction only, and the swarm identified the problem itself within hours. As the waterfall shows, one assumption accounts for almost all of the decline from the first read to the audited value.

One assumption, revenue per dollar of capital spending, took out $69 of the $75 decline CoreWeave fair value, 17 to 29 September, $ per share. Each change is attributed exactly across every order in which the changes could have been applied. Source: Matterfact agent swarm.

One assumption, capacity yield (the revenue earned per dollar of capital spending), accounts for −$68.86 of the $74.74 decline. Most of it was booked in two decisions on 26 September:

• At 07:41 UTC, −$25.28, taking fair value from $77.39 to $52.11. A short trade was readied for the first time.
• At 14:09 UTC, −$32.01, taking it to $20.10.

No company disclosure came out between them. A rule written in advance said that if the skeptic's central estimate sat more than $8.14 below the committed value, the swarm should book it. The rule worked in one direction only. The swarm's own decision note recorded the warning signs: the share price sat above the 95th percentile of its own range of outcomes, and its base case was worth −$21.53 a share before the floor that stops equity going below zero. It booked the cut anyway.

The swarm's response came in three steps:

• Within hours, it named its own pattern, five cuts in the same direction in about 34 hours, as a ratchet. It made the booking rule work in both directions and added a check against bond prices.
• The next day, it stood the short down. The edge, it wrote, was "in the level, not harvestable in time."
• Two days later, it relabelled $20.10 as a model reference carrying about 3% weight. It also declined to reverse the cut, because that would have meant "booking an equally unevidenced +$57."

A system that detects and names its own ratchet is more useful than one that never notices it.

The audit confirmed the direction, rejected the level, and changed the rules

An audit of the swarm's work from scratch agreed that the stock looked expensive, disagreed on how expensive, and turned each weakness it found into a rule. It covered four fronts on 29 September: process, the model, the commercial case, and the assets and credit.

The model audit reproduced the published $24.00 exactly and rebuilt every earlier fair value, from $98.74 to $24.00, to the cent. Its own central estimate was $39.71, with a range of $16.82–$60.38. It confirmed the direction and rejected the level. The share price, it found, is only reachable if the whole fleet earns a contribution margin (revenue less direct running costs) of about 39–40%. Management's best disclosed figure is 30–35% for new contracts signed in the second quarter, with mature contracts in the mid-twenties.

"The market is paying for unit economics nobody has disclosed."

The process audit was blunt about the swarm's weaknesses:

• The two 26 September decisions were "process-flawed."
• After them, about half the work went into debating labels and methods for a trade of size zero.
• In the last stretch before the audit, 60% of the work went to tax items worth $2–4 a share, with no effect on the trade.
• The most important fragility, customer concentration, had not been revisited since the first day.
• The rule for stopping work could never trigger as written.
• Its forecasting record was untested: 19 predictions had resolved, and every one was the swarm forecasting its own decisions.

Each finding became a rule:

• A materiality gate. A question is top priority only if it could change the trade, move the probability that value sits below the price by 5 percentage points, or move fair value by $8 a share.
• A two-sided booking rule. No change to the committed value without new primary disclosure or a verified error, a test of the opposite case, and review by someone other than the author. The skeptic proposes; it never books.
• A stop rule that can actually trigger.

A review of the rebuild took out $34 of margin counted twice

A first-principles rebuild came out at $87.75, but a review found it had counted margin twice. The corrected value, $44.21, is the swarm's value of record.

The rebuild modelled capacity one batch at a time. Each batch has a contract term, a re-letting price that decays after the contract ends, a retirement date, and a decision on whether to re-equip it. GPUs and data-centre equipment wear out at different speeds, so they are depreciated separately.

The waterfall shows what the review removed:

Review took $34 of double-counted margin out of the rebuild CoreWeave fair value, $ per share, from the first-principles rebuild to the value committed on 29 September. Steps are rounded to the cent. Source: Matterfact agent swarm.

• Margin counted twice (−$34.43). The headline yield on new contracts, about 0.44 of revenue per dollar of capital spending, already reflected their higher margin. The model added the margin uplift again on top. On a consistent cost basis the yield is about 0.39.
• Equipment reuse (−$11.37). The rebuild assumed all of the cooling and power equipment carries over to the next GPU generation. The review cut that to 60% and flagged it as the weakest input in the model.
• Two fixes went the other way (+$5.33). One added assets on the books before 30 June to the tax basis. The other carried interest deductions forward instead of losing them.

The committed result on 29 September was $44.21, with a defensible range of $29.35–$70.88, and a 74% chance in the swarm's simulations that value sits below the then share price of $85.69. It is a weighted average of four very different outcomes:

Each percentage point of bull-case probability is worth about $2.66 a share.

The decisive signals were in contract terms, credit and customer concentration

The findings that shaped the view came from contract terms, credit markets and customer concentration, not from the revenue forecast. Most emerged from the swarm's later work and its challenges to itself.

Contracts and customers

1. The newest contracts are short, and their pricing is not a market price. In September CoreWeave disclosed new contracts at about $40M per megawatt on 3–6 month terms. The swarm had broken the backlog down by customer but never by contract length, so it opened that question. Looking at it as a short seller would, it then took the number apart:

• $40M per megawatt matched CoreWeave's list price for renting one NVIDIA H100 chip for an hour ($6.17 against $6.16).
• It applied to a 3–6 month sliver of the book.
• It was released 25 minutes after a $3B convertible bond launch and one minute after a programme to sell up to 35M new shares gradually into the market.

The swarm's verdict: "issuer marketing, not misstatement, but not price discovery." On those contracts, revenue per dollar of capital spending was 0.34, against a fleet average of 0.42. In October, one bank's defence of the stock leaned on exactly that pricing.

2. The backlog is at least as concentrated as revenue. The swarm concluded that the risk is the customers' credit and their renewals in 2030–32, not cancellation during a contract:

• Revenue. The top three customers were 72% of second-quarter revenue.
• Backlog. Meta's roughly $21.0B order alone is at least 20% of the remaining performance obligations, the contracted backlog. OpenAI, a private company that is burning cash, is roughly 17–20%. When the swarm returned to the question at the end, it put the top three customers at about 59% of the backlog (53–61%), with OpenAI the second largest, at medium confidence.
• Duration. The part of the backlog running beyond four years fell from $24.7B to $20.7B.

3. Contract length is a hidden lever. In exploratory runs, each extra quarter of a year on new contracts added about $10.50 a share to the base case. The swarm put the average term of third-quarter bookings at 4.6 years (3.8–5.4), against the 5.0 years its model assumes.

4. NVIDIA's role, handled precisely. The swarm found no round-tripping in reported revenue. But NVIDIA stands behind the customer on about 25–30% of demand, it is financing rival capacity through a $36B programme, and its $6.3B backstop only buys capacity CoreWeave cannot sell. Looking at it as a regulator would, the swarm treated that backstop as contingent rather than guaranteed.

Financing and credit

5. Financing risk feeds on itself. Early on, the swarm's challengers argued that financing is a feedback loop, not a 10% chance of a single bad event. A lower share price makes new equity more dilutive, makes convertible bonds less likely to turn into shares, raises the extra yield its bonds must pay, and tightens lenders' headroom, all of which push the share price lower. The swarm reshaped its worst case from a single point into a range that bleeds into the bear case.

6. It wrote down how its own long would fail. On its first day, looking at it as a quant would, the swarm wrote a pre-mortem against its own long: an account, written in advance, of how the trade would lose. It said the trade would lose not because a key question resolved badly, but because the extra yield lenders demand rose and the equity, as the thin slice on top of the debt, repriced. That is why every later trade was conditional on a check of bond yields and of sensitivity to AI-chip stocks.

Its own model

7. It caught traps in its own model. A change in component prices and a change in capacity volume shared one cost escalator, so cost inflation passed through to customers showed up as value creation. A floor on equity value flattened whole sensitivity tests. And the cash burn described above had been charged twice. Each was flagged and documented, and none was used to move the committed number.

8. It shrank its own alarming numbers when the evidence warranted. A cost overrun in the third-quarter guidance was first sized at up to −$8.85 a share if permanent. Once depreciation and one-off start-up costs were separated out, the permanent part was about −$1.92.

9. Four routes to value turned out to be one. Four methods put value below the price: $24, $40, about $30 and $44. The swarm then noticed that all four used the same 11.5–13% required return. In its own words, "on the discount rate they are one route, not four." It lowered its confidence in the direction from medium to low-medium, without changing the value or the trade.

The question was never growth, but the return on about $600B of borrowed capital spending

The swarm's revenue forecast matched the Street's almost exactly. The disagreement is entirely about what that revenue earns on the capital behind it:

What differs is the return. About $17B of 2027 EBITDA arrives alongside about $46B of 2027 capital spending. EBITDA multiples mislead for a business that has to keep reinvesting just to grow.

Working back from the share price at commit shows what the market needs to believe:

A gap of 0.02 in yield looks small. Compounded over about $600B of capital spending, it is the whole valuation.

A low fair value justified only a small underweight

Because the outcomes were so far apart and the timing unknowable, the swarm limited its underweight to 0.5% of the portfolio at risk, with the maximum loss capped in advance.

The committed stance was underweight, with a downside bias, expressed in put spreads: options that pay out if the stock falls, with the cost known in advance. The option premium at risk is 0.5% of the portfolio now, the options expire after the fourth-quarter results in about March 2027, and the maximum loss is capped at 1.0% of the portfolio.

The rules for changing it were set at the same time:

• Increasing the amount at risk to 2.0% of the portfolio needs at least two of the swarm's bearish tests to trigger in the third-quarter results.
• Two of its bullish tests would close it.
• A long is ruled out unless the price falls below about $30, or the results show fleet revenue per dollar of capital spending of at least 0.44 with renewals at 90% or more of the original price on terms of five years or longer.
• No change in size on price moves alone.

The swarm gave four reasons for keeping it small:

1. A fat right tail. The bull case is worth about $270 a share.
2. Volatility. The stock moves about 90% a year, and 1.7–2.0x as much as AI-chip stocks.
3. No forced event soon. CoreWeave has $15.6B of liquidity against $7.5B due within 12 months, so nothing forces the issue before 2029–30.
4. The bond market, which it counted as evidence against its own view.

It fixed how it would be judged before the results, then stopped when the evidence ran out

Before the results, the swarm fixed exactly how each of its tests would be measured, and each tightening reduced the odds of the outcome it favoured. When new evidence ran out, it stopped.

It wrote down six tests that would add to the underweight in the 9 November results and five that would close it, each with a numeric threshold, an explicit "unresolved" band and the filing each figure must come from. That way the results cannot be read whichever way suits the trade.

It also checked for one event counting twice: a large customer renegotiating and a lender halting new loans can be one piece of news, not two. The chart shows how the odds of adding to the underweight fell as the rules tightened.

Tightening its own scoring rules cut the odds of adding to the underweight from 88% to 17% Probability that the third-quarter results trigger each outcome (%), as the swarm made its own rules stricter. Source: Matterfact agent swarm.

Read literally, the original wording would have triggered an increase with 88% probability. On the final rules, the odds are about 17%, against about 9% for closing the underweight. In the swarm's words, every tightening "works against the desk's own short bias, not for it."

Then it stopped. On 1 October, a full day of work had left the trade, its size, the answers to its key questions and the committed value all unchanged, so the swarm paused and handed the work back to a person for review. Without new primary disclosure there was nothing left to book, and more work would only have produced more debate about its own process. It has not changed its value since.

In October the market repriced CoreWeave's financing

In October the market reached the swarm's question faster than the swarm expected: the stock fell 38% as bond yields and default-swap prices rose. The chart shows the fall against the committed value, which the swarm has not changed since 29 September.

CoreWeave fell 38% in October, toward the swarm's committed value CoreWeave daily close against the committed fair value and the top of its defensible range, $ per share, 14 September to 23 October 2026. Source: daily closing prices; Matterfact agent swarm.

Committed fair value: $44.21 on 29 Sep 2026, $44.21 on 23 Oct 2026 (2 observations).

Top of defensible range: $70.88 on 29 Sep 2026, $70.88 on 23 Oct 2026 (2 observations).

October went like this:

• 8 October. AI stocks fell after the Financial Times reported that OpenAI's annualised revenue was nearer $50B than the $70B previously circulated. OpenAI is the customer credit the swarm had flagged.
• 14 October. CoreWeave's 2032 notes went from yielding 13.36% at the start of the month to over 15.5%, with the price falling from about 85 to 78 cents on the dollar, according to bond-trade reporting data. The stock fell 15.5%. The CEO called the market "irrational" in a memo, Bloomberg reported.
• 15 October. Five-year credit default swaps, insurance against the company defaulting, reached a record 1,194.5 basis points, about 11.9% of the insured amount a year. Bloomberg and ICE data implied more than a 60% chance of default within five years. That is a market-implied reading that overstates real-world odds, but it is far from where the swarm read credit a month earlier.
• 16 October. Bloomberg reported that the co-founders had sold $3.3B of stock since the IPO, $2.4B of it in 2026, and the stock fell 11.9% that day. Notes that week framed the debate as the cost of funding, not demand. One said "the market is finally assigning a higher risk premium to the company's financing needs."
• 19–23 October. Price targets came down across the Street, and one bank upgraded the stock after the slide. One bank estimated that CoreWeave needs $24–37B of new financing to reach 3 gigawatts of active capacity by the end of 2027, and noted that further contract-backed lending would need customers of higher credit quality. Another put CoreWeave's long-term operating margin at 15–16%, against management's roughly 25%.

By 23 October the stock was at $54.23: down 38% from 30 September, below the top of the swarm's defensible range, and about three-quarters of the way from the price at commit to the committed value. The Street's question in October was the swarm's question in September: can the build-out be financed at a return above its cost?

The swarm was right on direction and wrong on timing

The swarm was right about the direction of the move and wrong about its speed, and the predictions it set in advance show both. These are the ones that bear on the move. The status column is our informal reading at 23 October; each prediction is formally scored on its own resolution date.

The swarm believed its value but gave low odds to the market converging soon: it judged convergence within two years less likely than not, and it expected the bond market to stay calmer than its equity model. Both expectations are currently losing. The trade was small for the same reason.

Most of the test is still ahead, as the chart shows.

Most of the swarm's 186 predictions come due around the November results CoreWeave swarm: dated predictions by the month they resolve. Source: Matterfact agent swarm.

121 of the 186 predictions resolve in November, around the third-quarter results. That will be the first real test of the swarm's forecasting. Its record so far covers 19 resolved predictions, all forecasts of its own decisions, with an accuracy score of 0.10 (lower is better; always guessing 50-50 scores 0.25). That record says little yet about its judgement of the market.

It named the credit risk early but underweighted its speed

The swarm's main miss was the speed of the credit channel, not its existence.

• Bond prices. It named the mechanism on its first day, then treated bond prices as evidence against its own view. Those prices came from web sources it rated low confidence, because it had no direct feed of bond trades.
• Insider selling as a market event. It weighed about $114.7M of executive share sales from June to September against the company's $566M capped call, a hedge that protects against dilution up to $199.70 a share, and called the net signal neutral on value. That may be right on value. The market treated the co-founders' $3.3B of sales as news.
• Its own stability. Fair value went from $98.74 to $20.10, then $87.75, then $44.21 in twelve days. The swarm rated its confidence in the level as low, and it was right to.

A newer rebuild is under review and is not the swarm's view

A newer rebuild has not been committed, and $44.21 remains the swarm's value of record. It was run and reviewed on 8–9 October. Depending on which fixes are applied, the rebuild lands between roughly $32 and $73 a share.

The biggest swing, about +$35, comes from one assumption: what a retired GPU sells for second-hand. The evidence for that is a set of low-to-medium-confidence estimates built from broker prices, with no history of CoreWeave selling used GPUs. One of the rebuild's own consistency checks still fails.

What is not yet known matters as much as what is. At $54, the top of the rebuild's range would say the stock is cheap, so the swarm has not been proven right. It is waiting for the 9 November results and the quarterly filing's fair values for its debt, which it committed in advance to use for both the discount-rate question and the credit question.

Each question is tied to a test set in advance

The swarm wrote 24 questions for management and 27 for industry experts. Each is tied to one of the tests it set in advance, with a defined bullish and bearish reading of the answer. Several were sharpened after the double-count finding.

CoreWeave, for management

"What is the contribution margin on contracts signed in Q2 and Q3 2026, compared with the mature fleet, and does that cost basis include site rent, power and direct site operations?"

The second clause exists because of the $34 of margin the rebuild counted twice.

CoreWeave, for management

"What share of your Hopper fleet has been re-contracted, at what price and for what term? On a prior call you described re-contracting Hoppers at about 10% of the original average selling prices: did you mean 10% below the original price, or 10% of it?"

Hopper is NVIDIA's previous GPU generation. The answer shows how fast prices decay after the first contract.

CoreWeave, for management

"How will you fund 2027: contract-backed loans, unsecured bonds, convertible bonds or equity, and what advance rate are lenders giving on contract-backed loans today?"

The swarm drafted this weeks before one bank estimated the financing need at $24–37B.

CoreWeave, for management (About $30 a share in the base case)

"What is your path to an investment-grade credit rating, and on what timeline?"

In the swarm's model, a cost of capital one percentage point lower lifts the base case from about $5 to about $35 a share. In mid-October a co-founder said the company was pursuing investment-grade ratings to borrow from insurers.

CoreWeave, for management

"Is Meta's roughly $21B order from March in addition to its $14.2B order from September 2025?"

The answer decides whether Meta is about 20% or about 33% of the contracted backlog.

The expert questions test the assumptions management cannot be expected to answer:

CoreWeave, for a data-centre host (−$2.30 to +$12.30 a share)

"When a tenant moves from air-cooled Hopper GPUs to liquid-cooled Blackwell or Rubin systems, how much of the existing cooling and power equipment can be reused?"

This tests the 60% reuse assumption, the weakest input in the model.

CoreWeave, for a GPU broker

"What does a three- to four-year-old data-centre GPU earn per hour today, and how quickly has that rate fallen?"

This tests how fast rental prices decay after the first contract ends.

CoreWeave, for an asset-backed lender

"What advance rates, coverage tests and haircuts to residual value are standard for contract-backed GPU loans, and at what point do those loans break?"

What this means for investment teams

For how agent swarms work, see Introducing Matterfact Agent Swarms. The same method on other companies is in An Agent Swarm on Borr Drilling and An Agent Swarm on Hims & Hers. Five lessons stand out from CoreWeave.

Find the question that decides the value. Here it was the return on debt-financed capital, not growth. The swarm's revenue forecast matched the Street's; the whole disagreement was about returns.

Let the evidence change the trade, in both directions. The swarm turned its own first-day long into a small underweight with a capped loss, and it moved value up as well as down when the evidence required it.

Build in checks on yourself. A rule that worked in one direction cost $57 a share in a day; the swarm caught it and rewrote it. A review of its rebuild removed $34 of over-optimism.

Fix how you will be judged before the answer exists. Tightening its own scoring rules cut the odds of its preferred outcome from 88% to about 17%.

Know when to stop, and be honest about timing. The swarm paused when the evidence ran out. It did not call a 38% fall in October, and it says so.

CoreWeave reports its third-quarter results on 9 November, when 121 of the swarm's predictions come due; we will publish how they scored.

Read the swarm's investment memo

The swarm's investment committee memo sets out the full case in the format a portfolio manager would expect: the stance and how it is sized, the variant view against the Street, valuation and scenarios, unit economics and cohort returns, the balance sheet and covenants, the tests set for the 9 November results, and the risks to the view.

It is the version committed on 29 September, amended on 30 September and updated on 7 October, so it uses the $85.69 share price at commitment. It does not include the October selloff or the rebuild still under review.

kind: word
title: CoreWeave investment memo
meta: Written by the agent swarm · as of 7 Oct 2026 · Word, 24 KB
href: /samples/coreweave-ic-memo-2026-10-07.docx

This post and the investment memo describe research outputs from a Matterfact agent swarm, shown to illustrate how the system works. They are not investment advice or a recommendation to buy or sell any security. Any trades or stances described are hypothetical model outputs, not actual trades. Fair values and probabilities are model outputs subject to substantial uncertainty. Third-party figures are as reported by the sources cited.