F.F

F.F Research · AI · 28 August 2026

Who books GPU use rights as rent

The public U.S. sample of ASC 842 lessor operating leases on company-owned, identified GPUs is two names: Boost Run and Digi Power X.

Thesis · Public sample

One core 842 GPU lessor; one similar line

97.5%BRUN H1'26 lease / total
16.3%DGXX Q2 GPU 842 / total
$1.9BBRUN TCV · ~$0.4B powered
0%CRWV · NBIS · CHRN · WYFI · IREN 842 lessor

The split is not Kubernetes on the product page. It is whether the contract hands control of an identified asset. APLD / WULF 842 is halls and power. WhiteFiber 842 is the opposite direction: lessee GPUs sold as 606 cloud.

Article architecture

Test → Boost Run → Digi Power X

  1. 01

    ASC 842 vs ASC 606

    Two gates. Platform is a common reason, not the definition. How you sell is orthogonal to how you book.

  2. 02

    Boost Run

    The only listed core 842 GPU lessor. Finance-leased cards, colo halls, take-or-pay TCV waiting on rack-and-stack.

  3. 03

    Digi Power X

    Same accounting on one line. Variable rent, ATM cash, and a $1.1B Cerebras colo that is not NeoCloud.

Chapter 01

ASC 842 vs ASC 606

The same dedicated GPU contract can enter only one U.S. GAAP framework.

  1. 1.1 How each books revenue, and how much is 842
  2. 1.2 The legal tests
  3. 1.3 A cloud platform is a common reason, not the definition
  4. 1.4 What actually diverges on the statements

Chapter 1 · 1.1 Ranking

Product-layer height is not 842 purity

BRUN97.5% 842
DGXX16.3% GPU line
CRWV0% · ~95–98%+ as 606
CHRN0% · 6,144 H100s still 606
APLD/WULFhall 842 · no owned GPU-hours

Sorted by GPU / cloud product-layer share, not by 842 purity. BTDR is IFRS 15. WYFI GPU servers are a lessee cost.

Chapter 1 · 1.1 Direction

Almost every NeoCloud has 842 — as cost

Revenue-side 842

Customer money booked as operating-lease rent. Public sample: BRUN core, DGXX GPU line. Asset stays on the lessor’s books.

Cost-side 842

Halls, power, even GPUs leased in. WhiteFiber leases GPUs to sell 606 cloud. APLD / WULF lease out cabinets and power; the tenant brings chips.

Chapter 1 · 1.2 Legal test

Both gates must pass

Identified asset

Specific GPUs / servers. No substantive substitution.

Control of use

Customer directs workload, login, OS, utilization.

Lease

Fail either gate → 606 service.

Two sequential gates: identified asset with no substitution, then control of use. Both pass to a lease; either fail to a service.

Chapter 1 · 1.2 Critical audit matter

A CAM is judgment, not a qualified opinion

Deloitte marked certain CoreWeave customer contracts as a Critical Audit Matter: cloud service (606) or lease of cloud equipment (842). The company kept substitution and direction. The opinion on the statements is unqualified.

Nebius’s 20-F marks the same question. The second CAM layer is variable consideration and whether a reclassification slides into sales-type — GPU life 3–5 years, CRWV weighted term about 5 years.

Chapter 1 · 1.3 Definition

A platform is a common reason for 606, not the definition

  • Orchestration and multi-tenancy usually fail both gates — CoreWeave, Nebius, DigitalOcean.
  • Bare metal is not automatically 842. A control plane is not automatically 606.
  • DGXX’s website looks on-demand; the booked contract is 24-month exclusive bare metal, 842.
Left: identified exclusive servers under customer command, 842. Right: a substitutable pool dispatched by the vendor, 606.

Chapter 1 · 1.3 ChronoScale

Dedicated is not a lease and not take-or-pay

CoreWeave

Dedicated clusters, company keeps which machines. Take-or-pay committed. All 606. CAM.

ChronoScale

All ~6,144 H100s to Together AI. Unit price can change. 12 months, then 60-day rolling. Not take-or-pay. 606.

Boost Run

Specifically identified servers. Exclusive control. Non-cancellable GPU lease. Core 842.

Giving a whole cluster to one customer only says the product is dedicated. It does not imply a lease, and it does not imply take-or-pay.

Chapter 1 · 1.3 Software

Thin software can sit on identified GPUs

  • BRUN delivers through IaC and hosted Kubernetes / vCluster.
  • The quarterlies do not write Kubernetes as the dominant performance obligation.
  • Policy: if software becomes the thing being sold, the whole arrangement can move to 606.
A thin Kubernetes sheet sits on identified exclusive GPU servers; a customer hand reaches through to a circled machine.

Chapter 1 · 1.4 Orthogonal

How you sell is not how you book

Accounting → Contract economics →

Take-or-pay · 606

CRWV committed. Capacity SLA. Company keeps swap.

Take-or-pay · 842

BRUN core. Non-cancellable rent. Identified servers.

Usage · 606

CHRN. Unit price can change. Short cancellable term.

Usage · 842

DGXX GPU line. All variable. No material fixed minimum.

Chapter 1 · 1.4 Visibility

Four leftover-money rulers are not interchangeable

Ruler Who requires it What enters Example
RPO606 mandatoryUnsatisfied PO after constraintCRWV Q2 ~$103.7B
BacklogNobodyRPO + undelivered commitmentsCRWV Q2 $104.2B
TCVNobodySigned face value, even unpoweredBRUN $1.9B · ~$0.4B live
Lease payments842 footnoteFuture fixed rent onlyDGXX GPU footnote almost empty

ARR is a flow. These four are stocks. BRUN mid-year ARR $145M is not TCV. EV / RPO cannot be pasted onto TCV.

Chapter 1 · 1.4 Choice

Control in the contract chooses the standard

Run 606

Keep swap, dispatch, refresh. Combine compute, support, software into one PO. Mandatory RPO. Classification is watched; sales-type is a left tail.

Run 842 operating lease

Legal story matches identified machines. Fixed-rent receipts table is harder than TCV. Substitution is given away. Software dominance can move the whole book to 606.

Chapter 02

Boost Run

The cleanest listed bare-metal 842 GPU lessor. Halls and cards do not start on the same day.

  1. 2.1–2.3 Business, Blue Fire lineage, controlled-company board
  2. 2.4–2.5 Finance-leased GPUs, colo operating-lease ROU
  3. 2.6–2.8 TCV, statements, delivery bottleneck

Chapter 2 · 2.1 Purity

Lease share rose because scale drowned mining

FY2578.9% · $21.2M
Q1'2696.9% · $10.6M
Q2'2697.7% · $30.4M
H1'2697.5% · $41.0M

H1 blockchain about 2.5%. IaC and vCluster exist; Kubernetes is not a separate revenue line. Exclusive control of specifically identified GPU servers.

Chapter 2 · 2.2–2.3 Lineage

Blue Fire ops DNA, then a de-SPAC

Blue Fire

13 DCs · 7 countries · HFT credit

Boost Run

Karos self-funded LLC

Dell $1.44B

Five-year purchase + DFS

BRUN

Nasdaq listing

0 independents

Controlled company · ~90% vote

Chapter 2 · 2.4–2.5 Sandwich

Lessor to the customer; lessee twice

$328.5MFinance ROU · GPUs in
$122.0MOperating ROU · halls in
$54.9MOwned equipment · $28M idle

An operating-lease lessor does not create a lessor ROU. USD 122M is colo, not GPU income. Power and excess usage are variable and stay off ROU.

Boost Run as lessor to the customer and lessee to OEM finance and colo landlords.

Chapter 2 · 2.4 Fleet on the books

The live fleet is leased in, not cash-bought

Finance ROU$328.5M
Finance liab.$238.1M
Owned equip.$54.9M
H1 cash capex$29.6M vs $268M new ROU

H1 34 new finance leases, undiscounted $295.9M. 2027–28 about $100M a year of repayments. Dell 1.44B essentially fully allocated; 4–5B more in talks.

Chapter 2 · 2.6 TCV

$1.9B TCV is not $1.9B of live rent

YE25~$120M
Listing~$940M
Q2 TCV$1.9B
Powered~$0.4B

Management path: Q3 $0.8B live → Q4 $1.2B → 2027 Q1 $1.9B fully live. Deck path, not a 10-Q commitment. Average term ~3 years, average prepay ~22%.

Chapter 2 · 2.6 Contracts

TML is take-or-pay; Fluidstack is thinner

Thinking Machines Lab · $471.7M

5,000 B300 · 36 months · pay all fees regardless of usage. MSA filed. $3.59 / GPU-hour. Customer A $208M starts Nov-26; Customer B $223M starts Mar-27.

Fluidstack · $127M

Two years, pre-listing. Release writes inference as usage-based. No MSA exhibit. Cloud-platform wholesale, not a lab using the cluster itself.

Announced Q2 names sum to $679M, or $902M with subsequent-event Customer B. Management said “over a billion.” That increment is TCV, not an 8-K roll-up.

Chapter 2 · 2.7 Presentation

Reported gross margin is a line choice, not 842

BRUN FY2519.5%
BRUN Q1'2643%
BRUN Q2'2634%
CRWV FY2522% · inside CoR

Add Q2 colo $10.6M back to CoR and BRUN is ~40% of revenue, near CRWV 34%. Add GPU D&A and operations already lose. OCF $114M is customer deposits, not profit. Tax $55.7M is C-corp deferred, not cash.

Chapter 2 · 2.8 Bottleneck

Halls are waiting. The constraint is B300s

$3.59TML B300 long-contract $/hr
$2.21M6/30 fixed colo / month
$0.53MRock Island empty-rent floor
5–7 MWPaid, empty, waiting for cards

Two more 7-year halls, undiscounted $42.8M + $184.0M, target Q3 commencement — numerator steps up again. Demand is not the constraint. Rack-and-stack is.

Chapter 03

Digi Power X

Similar accounting on one line. Different firm: miner conversion, ATM cash, and a colo mega-contract.

  1. 3.1 Four lines in parallel
  2. 3.2 GPU 842 policy, SubQ, cash-bought fleet
  3. 3.3–3.4 Cerebras colo and ATM residual cash

Chapter 3 · 3.1 Mix

GPU 842 is 16.3% of Q2 — and will be diluted

Colo 60651.7% · $3.43M
Energy 60629.5% · $1.96M
GPU 84216.3% · $1.08M
Mining 6062.4% · $0.16M

H1 GPU / total 7.4%. Tier III AI segment “has not yet begun commercial operations” — Cerebras hall still under construction. Booked GPU rent already sits in that column.

Chapter 3 · 3.2 SubQ

Same 842 gates; no take-or-pay floor

DGXX GPU line

SubQ only booked customer. 24 months, expected TCV $19.6M, 15% prepaid. Consideration all usage-based. 842 footnote almost empty. Cash-bought PPE $48.3M. ~0.6 MW.

BRUN core

Fixed non-cancellable rent. Remaining-lease-payments table. Finance-leased fleet. Customer deposits $128.4M. Policy: certain contracts are 842 — website on-demand is not yet a 606 line, but watch Q3.

Chapter 3 · 3.3 Cerebras

$1.1B is colo, not NeoCloud, and not booked

40 MWIT · 10 years · ~$1.1B
$195/kWTake-or-pay · +3%/yr
$94MFull-fit year-1 run-rate
0Booked to date

Tenant brings chips. Phase 1 15 MW target 2026-12-15; Phase 2 +25 MW subject to financing. Isomorphic to WULF / APLD, not to BRUN. Once booked, the group looks more like a small-cap AI colo.

Chapter 3 · 3.4 Cash

$128.1M cash is an ATM residual

ATM net$248.6M
Cash 6/30$128.1M
H1 PPE$96.3M
H1 OCF−$10.6M

YE24 cash $1.70M. Interest-bearing debt near 0. Alabama 40 MW is not $128M of cash. CEO: ATM is a “necessary evil” to qualify for project debt. Other registered offerings add ~$16M besides ATM.

Conclusion

Two names. One test. Orthogonal economics.

  • Identified asset + control of use is necessary and sufficient. Bare metal and Kubernetes are neither.
  • Take-or-pay can be 606 (CRWV) or 842 (BRUN). Variable 842 (DGXX) has no floor.
  • RPO, backlog, TCV, and remaining lease payments are four leftover-money rulers.
  • BRUN’s constraint is cards into already-reserved halls. DGXX’s cash is equity residual; Cerebras will re-label the group.

Independent research, not investment advice. Figures from 10-Q / 10-K / 8-K / IR materials. Cutoff: 28 August 2026.