How is the ranking built?

Three axes, each scored 1 to 10:

Survivability. Balance-sheet strength, cash generation, funding access, and maturity profile, if capital markets closed for a year.

Vulnerability (10 = most exposed). Single points of failure: customers, counterparties, refinancing, regulation, dilution.

Growth potential. Backlog relative to size, revenue slope, and pipeline.

The combined score weights survivability at 40%, growth at 35%, and safety (10 minus vulnerability) at 25%.

RankCompanySurvivabilityVulnerabilityGrowthCombined
1Nebius (NBIS)75107.6
2Akamai (AKAM)9357.1
3IREN6796.3
4Hut 8 (HUT)6786.0
5Applied Digital (APLD)5785.6
6CoreWeave (CRWV)4895.3
7Cipher Digital (CIFR)4884.9
8TeraWulf (WULF)3973.9
Grouped bar chart of survivability, growth and safety scores for the eight companies with combined scores
The two quality poles are visible at a glance: Akamai owns survivability and safety, Nebius owns growth, and no one else leads an axis. Chart: DalalBytes, from the scores above.

This is a quality ranking, not a return forecast. The distinction matters twice in this list.

TeraWulf ranks last on structure and was still our Selective Positive on October 15, because its price had already paid for the structure. CoreWeave ranks sixth despite the largest backlog in the group, because the credit market is pricing its capital structure honestly.

Change the weights and the middle shuffles. The top two and the bottom one do not move.

Who are the eight?

Four species sell into the same shortage. Two GPU cloud operators (CoreWeave, Nebius) rent compute by the hour and carry GPU fleets on their own balance sheets. Three landlords (Applied Digital, Cipher, TeraWulf) sell powered buildings on 10 to 25-year leases. IREN is migrating from bitcoin mining to the operator model. Hut 8 is a developer-landlord that still mines: it builds gigawatt campuses, finances them like an infrastructure sponsor, and lets its compute segment, mostly American Bitcoin, pay today's bills. Akamai is an incumbent cloud and security company buying its way in with one very large capacity deal.

Oct 15, 2026CRWVNBISIRENAKAMWULFCIFRAPLDHUT
Price$69.69$188.41$30.39$92.32$12.62$10.04$20.39$76.09
Market cap~$38.4B~$47.7B~$12.0B~$13.3B~$6.3B~$4.2B~$6.2B~$9.4B
52-week range60.55-150.4073.52-299.8628.93-76.8770.82-165.4510.47-29.848.79-30.1418.36-50.7231.67-140.80
Anchor customersMicrosoft, OpenAI, MetaMicrosoft, MetaMicrosoft, NVIDIAThousands; Anthropic newAnthropic, FluidstackAWS, FluidstackCoreWeave; one IG hyperscalerOne unnamed IG tenant; Fluidstack (Anthropic)
Short interest (float)~16%~18%~22%Low~28%~22%High~15%

How do revenues and cash flows compare?

CoreWeave is the revenue king: fiscal 2025 revenue $5.13B, Q2 2026 $2.58B (+112% YoY), fiscal 2026 guided to $12.4-13.2B.

Nebius has the steepest slope: Q2 2026 revenue $582M (+454% YoY), ARR from $1.9B in March to $3.0B in June, guided to $7-9B by year-end.

Akamai has the largest durable base and the slowest growth: $4.21B in fiscal 2025, +5-6% in 2026, with Security +10%, cloud infrastructure (CIS) +39%, and Delivery -6%.

The landlords still report rounding errors against their contracts. Applied Digital booked $342M GAAP in fiscal Q1 2027 (+322%, mostly fit-out and hosting). IREN booked $707M in fiscal 2026, with AI cloud ($129M) passing mining for the first time. Cipher booked $60M of mining-only revenue in H1 2026. TeraWulf booked $45M in Q2 2026. Hut 8 booked $235M in fiscal 2025 and $74.9M in Q2 2026 (+81% YoY), almost all of it Compute, and its fiscal 2025 net loss of $248M was mostly a $220M unrealized Bitcoin mark. Its AI campuses contribute no revenue until 2027.

Cash flow separates the models more sharply than revenue does:

CompanyOperating cash flowFree cash flow character
AKAMFY2025 $1.52BFCF about $1.0B in FY2025; the only self-funding engine here
NBISQ2 2026 $2.25BPrepayment-driven; H1 capex $8.1B against $4.5B of OCF
IRENFY2026 $2.10BMostly a $1.7B deferred-revenue build in one quarter; ex-prepayment Q4 OCF about $89M
CRWVQ1 2026 $2.98BQ2 2026 FCF about -$5.7B; capex guide raised to $35-39B
APLDQ1 FY27 +$64MCapex of $2.07B in the same quarter
HUTH1 2026 -$32.8MMining cash flow is thin against the build; construction is pre-funded by non-recourse IG project notes, not by operations
CIFRQ2 2026 -$244MNegative in 7 of the last 8 quarters
WULFFY2025 -$123MNegative through the build-out

Prepayments are a genuinely better funding model than debt. They are also a promise: when prepayments slow, the capex commitments they funded do not slow with them.

What do the debt stacks say?

CompanyTotal debtCashSafety note
NBIS$10.1B$8.0BNet debt ~$2.0B, below one year of guided revenue
IREN$8.0B principal$5.9B ($7.6B w/ restricted)Investment-grade GPU financing near 6% for Microsoft; converts underlie up to 166M shares
HUT~$7.6B consolidated$0.2B unrestricted, ~$7B restricted$7.5B of the debt is fully amortizing, non-recourse IG project notes (BBB- and Baa2) for River Bend and Beacon Point; no recourse debt at the parent after the Coatue convert converted; Bitcoin holdings sit inside an $8.1B combined liquidity figure
AKAM$7.6B converts$4.6BNet ~$3.0B; the only investment-grade balance sheet (Baa2 / BBB, negative outlook)
APLD$6.4B$3.7B incl. restrictedProject debt termed to 2031+; about $13B of further funding needed (JPMorgan estimate)
CIFR$6.0B principal$0.8B + $3.2B restrictedEquity $562M; company net debt ~$1.5B when restricted cash is netted
WULF$5.7B face~$3.0B incl. restrictedBook equity $147M; a ~$3.5B raise still ahead
CRWV$35.1B balance sheet$5.5BEquity ~$5.0B; bonds at 13-15% yields; CDS ~1,000bp; rated B+ / BB-
Bar chart of total debt versus cash for the eight companies
Seven of the eight carry debt and cash in the same ballpark. CoreWeave is the outlier: roughly $35B of balance-sheet debt against $5.5B of cash. Chart: DalalBytes, from latest reported balance sheets.
Reuters (October 15, 2026) puts CoreWeave's total obligations near $50B including later raises and leases, and LSEG data implies a roughly 63% five-year default probability from its CDS. Cantor's bull case (2027 capacity nearly sold out; NVIDIA's $6.3B unsold-capacity backstop) and the CDS can both be right. They are pricing different years.

How does the technology actually differ?

CompanyPlatformScale nowTechnical differentiator
CRWVGPU cloud operator1.5 GW active, ~4.2 GW contracted; 250,000+ GPUs, 51 data centersGB200 and GB300 NVL72 at scale; Vera Rubin NVL72 in production from Sep 2026; SUNK, Kubernetes service, Mission Control; Weights & Biases; ClusterMAX Platinum three straight times
NBISFull-stack GPU cloud~220 MW end-2025, guided 800 MW-1 GW end-2026Own data centers and cloud platform; Token Factory inference; NVIDIA design partner (9.3% holder); 41% adjusted EBITDA margin at small scale
IRENAI cloud, ex-miner~40 MW AI live (Jun 2026); 0.3 GW targeted in 2026; 750 MW Childress campus, 2 GW SweetwaterNVIDIA DSX AI-factory reference design toward 5 GW; ~150,000 GPU target; vertically integrated power and sites
AKAMDistributed edge cloud4,300+ points of presence, 130+ countriesCPU-first inference at the edge plus GPU expansion; Guardicore segmentation and API security monetize the defense side of the same build-out
APLDDeveloper-landlord250 MW live at Polaris Forge 1; 1.41 GW leasedLiquid-cooled, high-density campus design; ChronoScale GPU-cloud arm (GB300 NVL72, liquid-cooled)
CIFRLandlord (ERCOT)907 MW operating plus contractedAir and liquid cooling at Black Pearl; ERCOT interconnection queue positions; up to 2.5 GW of behind-the-meter gas in development
HUTDeveloper-landlord plus ASIC compute949 MW of AI capacity contracted, 1,330 MW of utility capacity under construction; ~700 MW of mining infrastructure serving American BitcoinPower-first site origination and an IG project-finance template: River Bend priced at a reported 95% loan-to-cost, Beacon Point 20bp inside it; an NVIDIA DSX redesign lifted Beacon Point IT capacity 57% inside the same footprint
WULFLandlord~102 MW earning rent; ~336 MW under constructionLiquid-cooled colocation buildings; ~93% zero-carbon power mix (company figure); demand-response grid flexibility

CoreWeave's technology risk is accounting, not capability. The fleet is depreciated over about 6 years while critics argue the economic life is 2-3. Q2 2026 depreciation of $1.39B already runs at 54% of revenue. If the critics are right, reported margins overstate economic margins across every operator using long lives; CoreWeave is simply the largest case.

The landlords' technology is power engineering. Liquid cooling density, interconnection queue position, and delivery speed decide who gets the next lease. In a shortage, interconnection rights are the technology.

Akamai is the only one selling the other side of the trade. Every GPU cluster the others build expands the attack surface its security products defend, and its Anthropic deal rents CPU capacity at the edge rather than GPU capacity in a mega-campus. Different physics, different risk.

What does the deal ledger show?

First, the operators and the incumbent:

CompanyCounterpartyScaleTermValueCredit support and status
CRWVMicrosoftCore GPU capacityMulti-year~67% of FY2025 revenueAnchor tenant; concentration is the story
CRWVOpenAICapacity tranchesThrough 2029-31~$22.4B total (11.9 + 4.0 + 6.5)Includes a $350M OpenAI equity stake in CRWV
CRWVMetaCapacityTo Dec 2031; expansion to 2032$14.2B + $21B expansion (Apr 2026)Signed; expanding
CRWVAnthropicCapacityMulti-year (Apr 2026)UndisclosedSigned
CRWVNVIDIAUnsold capacity purchaseThrough 2032$6.3B backstopPlus a further $2B NVIDIA equity investment (Q1 2026)
NBISMicrosoft~300 MW, Vineland NJThrough 2031~$17.4B (to ~$19.4B with services)All contracted tranches delivered (Oct 2026)
NBISMetaDedicated capacity5 years; #2 from early 2027~$3B + $12B, plus up to $15B contingentTwo deals (Nov 2025, Mar 2026)
NBISNVIDIAStrategic investmentn/a$2B equity plus warrantsNVIDIA holds 9.3% (13G, Jul 2026)
IRENMicrosoft200 MW (Horizon 1-4), GB3005-yr average, through 2031~$9.7B20% prepay per tranche; Horizon 1 accepted, 3-4 due Q4 2026; termination rights on missed dates; $3.6B IG GPU financing at ~6%
IRENNVIDIA~60 MW cloud + DSX partnership5 years; partnership to 5 GW~$3.4B cloud dealPlus NVIDIA's right to buy 30M shares at $70
IRENFrontier AI lab (unnamed)AI cloudMulti-year (Aug 2026)UndisclosedReported to be Anthropic (The Information); unconfirmed
AKAMAnthropicDedicated CPU capacity (Project Plans 2-3)7 years~$11.6B, plus up to $9B expansionDelivery and SLA conditional; revenue H2 2027 to a ~$1.7B pace by end-2028; warrant for ~7.74M shares (~5%) at $111.33, vesting in tranches
AKAMUS technology companyRobotics development capacity4 yearsOver $600MInside $2.8B+ of CIS commitments signed year-to-date through Q2 2026

Then the landlords, where the lease is the product:

CompanyCounterpartyScaleTermValueCredit support and status
APLDCoreWeave (Polaris Forge 1)400 MW, Ellendale ND15-year leasesPart of ~$36B base total250 MW live; Building 2 Phase 2 delivered Oct 1, 2026; Building 3 building
APLDIG hyperscaler, unnamedPF2, PF3, Delta Forge 1 and 2 (DF2: 210 MW)15 years (DF2)DF2 ~$5.2B (Jun 2026)Same tenant at four campuses; still unnamed at Oct 7, 2026
CIFRFluidstack (Barber Lake)207 MW criticalExtended to 20 years (Sep 2026)Over $9B at the siteGoogle backstop $1.73B; Google warrants 24.18M shares at $0.01; rent phases Q4 2026-Q1 2027; CIFR absorbs $359.3M of first-loss overruns
CIFRAI lab, unnamedBarber Lake, second term10 years after the Fluidstack term~$5.2B incrementalBinding commitment on substantially consistent economics (Sep 2026)
CIFRAWS (Black Pearl)~300 MW15 years~$5.5BDirect lease with AWS; rent commenced Aug 2026, two months early
CIFRIG hyperscaler, unnamed (Stingray)Third campus15 yearsUndisclosedFinanced by $810M Stingray Compute notes at 6.000%
HUTFluidstack (River Bend, LA)245 MW15 years$7.0B base (to ~$17.7B with renewals)Google backstops base-term lease payments; Anthropic is the workload; first data hall Q2 2027; financed with $3.25B of BBB- IG notes, non-recourse
HUTIG tenant, unnamed (Beacon Point 1, TX)352 MW15 years$9.8B baseTriple-net, take-or-pay; reported to be NVIDIA leasing for Anthropic's Lambda capacity (Bloomberg), unconfirmed by Hut 8; financed with $4.25B of Baa2 notes; delivery from Q3 2027
HUTSame tenant (Beacon Point 2, TX)352 MW15 years$9.8B base (campus $19.6B; $50.2B with renewals)Signed Jul 2026, five months after Phase 1; Phase 2 financing not yet closed; delivery from Q2 2028
WULFAnthropic (Justified, KY)~401 MW20 years, two 5-yr extensions~$19B (to ~$33B with extensions)Direct lease; credit supporter unnamed; delivery H2 2027 to early 2028
WULFFluidstack (Lake Mariner)~360 MW (CB-3/4/5)10 years~$6.7B initial (~$16B with extensions)Google backstop ~$3.2B via recognition agreements; Google warrants 73.58M shares at $0.01
WULFCore42 (Lake Mariner)~60-72.5 MW10 years + renewals~$1.1B (~$2.5B with extensions)G42 guaranty; $90M prepaid rent; fully commenced by Mar 2026

Read the credit-support column, not the value column. The deals with named, rated money behind them (Microsoft prepayments, AWS direct, Google backstops, G42 guaranty) behave like financing. The deals with unnamed support behave like hope, and the market prices them accordingly.

Whose moat is real?

NBIS: platform plus patron. NVIDIA equity, hyperscaler prepayments, and a full-stack cloud create switching costs. Real, but partly rented from NVIDIA's goodwill and allocation.

AKAM: the only compounding moat. A two-decade edge network, a security franchise growing double digits, thousands of diversified customers, and now an $11.6B Anthropic anchor. Slow, wide, durable.

IREN: power plus partnership. More than 5 GW of grid-secured pipeline and NVIDIA's DSX imprimatur. The mining moat ends by design, so this one must be re-earned in cloud operations.

CRWV: scale and software, negative financial moat. The best operator technology in the group, wrapped in the worst funding structure. It compounds only if capital stays available.

APLD: delivery reputation. Leased 1.41 GW across five campuses to two counterparties. One unnamed tenant deep at four of them.

HUT: the financing template. Hut 8 was the first sponsor to fund a single-asset AI campus in the investment-grade bond market, then repeated the trade weeks later on better terms, and it holds an 8.7 GW power-first pipeline behind the two financed campuses. Like WULF's Google structure, the template will diffuse. What does not diffuse as quickly is the combination: IG-funded campuses, a Bitcoin treasury, and a majority-owned miner (about 61% of American Bitcoin) paying cash while the rent is still two years away.

CIFR: ERCOT positions plus the AWS proof. The pipeline is large (5.3 GW) and mostly uncontracted.

WULF: site scarcity plus structure invention. The Google recognition agreement was genuinely innovative and is now industry-standard, which is another way of saying the moat diffused.

How big are the pipelines against the companies?

Bar chart comparing contracted backlog headlines with market capitalizations for the eight companies
Every landlord trades at a fraction of its contracted headline, which is either the opportunity or the warning, depending on whether the funding arrives. Bases differ by company; see the footnote on the chart. Chart: DalalBytes.

Nebius's pipeline is the most de-risked relative to its size: tranches delivered, prepayments covering 50-60% of deal capex, paybacks near 22 months.

CoreWeave's is the largest and the most concentrated: its top three customers plausibly exceed 85% of forward revenue.

Akamai's is the smallest relative to its revenue, and the only one that is upside rather than survival.

Hut 8's contracted book is the largest relative to its size after Nebius's: $26.6B of base-term leases against a ~$9.4B market cap, with an 8.7 GW pipeline behind it.

TeraWulf's, Cipher's and Hut 8's convert to rent latest (2027-28 weighted), which is why their equities are the most violent in both directions.

Why these ranks, axis by axis?

Survivability: AKAM (9) > NBIS (7) > IREN (6) and Hut 8 (6) > APLD (5) > CRWV (4) and CIFR (4) > WULF (3).

Akamai funds itself and answers to thousands of customers. Nebius holds $8.0B of cash against ~$2.0B of net debt. IREN's cash is matched to contracted GPU financing, but it carries $13.8B of capital commitments and a $25-30B fiscal 2027 capex guide that assumes open markets. Applied Digital termed its debt out and still needs roughly $13B more. Hut 8 reaches IREN's tier by structure rather than scale: no recourse debt at the parent, both campuses under construction pre-funded with fully amortizing IG notes, and Bitcoin inside the liquidity, set against only ~$0.2B of unrestricted cash and a revenue base that is still mining. CoreWeave and Cipher carry credit-priced leverage on thin equity. TeraWulf has $147M of book equity and a $3.5B raise ahead of it.

Vulnerability: WULF (9) > CRWV (8) and CIFR (8) > APLD (7), IREN (7) and Hut 8 (7) > NBIS (5) > AKAM (3).

TeraWulf concentrates about 48% of its contracted megawatts in one lease whose credit supporter is unnamed, inside a New York regulatory fight. CoreWeave is one Microsoft decision and one credit-market mood from a refinancing problem. Cipher is pre-revenue in HPC until Q4 2026 and absorbed $359.3M of construction overruns. Hut 8 parks 704 of its 949 contracted MW with one unnamed tenant, and Bitcoin mark-to-market swung its quarterly result by about $356M between Q2 2025 and Q2 2026, from a $217.6M gain to a $138.6M loss. Akamai can lose the entire Anthropic bet and remain a profitable incumbent; no one else here can say that about its anchor deal.

Growth: NBIS (10) > CRWV (9) and IREN (9) > APLD (8), CIFR (8) and Hut 8 (8) > WULF (7) > AKAM (5).

Nebius pairs the steepest slope with the largest backlog relative to its size. CoreWeave's growth is real and bought at the highest cost of capital in the group. IREN's contracted ARR of $4B stands against $707M of fiscal 2026 revenue. Akamai grows slowest in percentage terms and adds the most absolute profit per point of growth. Hut 8's contracted book is the biggest relative to its size in the group, but none of it earns before Q2 2027, which caps the slope.

What would change the ranks?

The calendar first: Nebius has not yet announced its Q3 date (third-party calendars split between October 22 and November 10, 2026; ARR against the $7-9B year-end guide is the checkpoint whenever it lands). Cipher reports around November 2 (its first HPC revenue). Hut 8 in early November (third-party calendars split between November 3 and 4, both estimated; the company has not confirmed a date). Akamai on November 5 (CIS acceleration, Anthropic capex pace). TeraWulf on November 6 (rent commencements, financing signals). CoreWeave on November 9-10 (backlog conversion, credit commentary).

Then the structural triggers:

Naming the Anthropic credit support behind the TeraWulf lease (and confirming the Cipher lab) upgrades both landlords.

A CoreWeave refinancing at single-digit yields upgrades it two places. A punitive one confirms the CDS.

Microsoft accepting IREN's Horizons 3-4 in Q4 2026 settles its delivery question.

Applied Digital naming its hyperscaler removes its largest valuation discount.

Nebius holding prepayments above 50% of new-deal capex keeps its funding model, and its rank, intact.

Hut 8 closing Beacon Point Phase 2 financing at or inside Phase 1 pricing proves the IG template repeats a third time. A wide or pulled deal would mark the template, and the rank, as a two-time trick.

Own the trade in layers: Nebius for slope with a funding model that works, Akamai for ballast that pays you to wait, IREN for torque with delivery risk, Hut 8 for the group's cleanest funding structure at a miner's earnings multiple, and treat the landlords and CoreWeave as what they are, levered claims on 2027-28, sized by their credit spreads rather than their backlogs. The backlog is the advertisement. The funding is the business.

Download the detailed comparison (PDF)

Frequently asked questions

Which AI infrastructure stock is safest?

Akamai, and it is not close. It is the only self-funding, diversified, investment-grade name in the group: $4.2B of revenue, about $1.0B of free cash flow in fiscal 2025, thousands of customers, and an AI capacity deal ($11.6B with Anthropic) that is upside rather than survival. It ranks first on survivability (9/10) and last on vulnerability (3/10). It also grows slowest, which is why it ranks second overall behind Nebius rather than first.

Why does CoreWeave rank only sixth despite a $104B backlog?

Because the ranking prices the funding, not the backlog. CoreWeave carries about $35B of balance-sheet debt (Reuters puts obligations near $50B including later raises and leases) against roughly $5B of equity; interest runs near 25% of revenue; its unsecured bonds yield 13-15% and its CDS near 1,000bp implies a roughly 63% five-year default probability (LSEG via Reuters, October 15, 2026). Its bulls (average target $137-141 versus a $69.69 close) are pricing 2027 capacity scarcity, and they may be right, but on survivability and vulnerability the structure ranks where the creditors put it.

Why does TeraWulf rank last here but get a Selective Positive in its own report?

Different questions. The ranking grades structure: $5.7B of debt on $147M of book equity, an Anthropic lease whose credit supporter is unnamed, and a $3.5B raise still ahead make WULF the most vulnerable balance sheet in the group. The report grades price: after a 61% drawdown, the equity trades near replacement cost for its contracted megawatts and below Morgan Stanley's bear case. A fragile structure can be a good trade at the right price; this ranking deliberately does not confuse the two.

Disclosures

Research and opinion, not investment advice. The author holds long positions in Nebius, IREN, and Akamai. The author holds no position in CoreWeave, TeraWulf, Cipher Digital, Applied Digital, or Hut 8 at publication. Positions may change without notice. Figures are drawn from public filings and reporting believed reliable but not guaranteed; verify before acting. DalalBytes research on individual names in this comparison: Nebius, Akamai, Applied Digital, TeraWulf.

© 2026 DalalBytes Research · For educational purposes only. Not investment advice.