Yantra · decision intelligence for AI factories

WHY NOWCompute is the moat. Operating it is the margin.

“The infrastructure layer is the current beneficiary of spend… it's the people building compute who are doing the spending. It continues to become harder to deploy compute as the demand curve goes vertical.” Nikesh Arora, CEO Palo Alto Networks, Jul 31 2026. When deployment is the bottleneck, the operator who converts megawatts into sellable, reliable GPUs fastest wins the cohort. That conversion (burn-in, health, topology truth, power) is a data problem. Yantra is that data layer.

A week later the framing hardened: “Power is THE binding constraint… not fanciful plans for power, future forecasts of BTM or distributed batteries — but energized power today,” with neoclouds judged “solely measured by energized compute online today.” Chamath Palihapitiya, Aug 7 2026. If financing turns on energized compute online today, then megawatt-to-sellable-GPU stops being an efficiency metric and becomes a financing one. That is the number every layer below moves.

Five days after that, an operator put a price on it. Nebius's Q2 2026 shareholder letter publishes annual contract value per megawatt as a management metric, reports payback falling to 1 year 10 months from a historical two-to-three years, and raises the year-end contracted-power target to 5 GW. Three opinions became one disclosed number. Everything below is a lever on it.

Contracted ACV per MW · Nebius, disclosed Aug 12 2026

Contracted value on deals signed, much of it against capacity not yet built. It is not realized revenue over active megawatts. The tile above is that, and the two do not belong on one axis.

Contracted power target, same company, five raises

Power keeps arriving faster than the plan. The conversion from contracted megawatt to sellable GPU is what has to keep up, and it is the part no procurement contract buys for you.

What a repriced megawatt does to the pitch

×3.7

A week of bring-up, repriced

One week of a 10 MW hall carries ~$2.3M of contract value at the $12M base, ~$3.8M at the Q2 rate and ~$8.7M at the short-term rate. Same week, same work. Yantra's velocity argument repriced with the market and no claim on this page had to change.

1 yr 10 mo

Payback is the CFO's version of this page

Payback is capex, ACV per MW, time-to-revenue and utilization. Yantra moves three of the four. Talk stranded GPUs with the VP of infrastructure; talk months off payback with the CFO.

SOFR+250

Lenders now underwrite working GPUs

The first ~$775M facility is secured on deployed GPU infrastructure and contracted cash flows. Fleet health and delivery-against-contract became collateral quality, and the evidence layer became a treasurer's problem.

4 → 5 yr

Useful life is an evidence argument

Server life was extended a year on "usage patterns and current utilization commitments." Failure rates, thermal history and real duty cycles are what an auditor asks for. Forge and Sentinel already collect them.

BEFORE LAYER 01 The clock starts years before the megawatt Where the 34 days sits · and the search problem underneath it

$6.7 trillion of capital expenditure lands in data centers through 2030, and almost none of that clock is measured in days. Site due diligence through permit-ready design runs 10 to 12 months. The power behind the site is a multi-year queue. Yantra owns the last 34 days. That is the point rather than a limitation: those 34 days are the only stretch an operator can still compress this quarter, and at Q2 2026 contract rates a week of a 10 MW hall carries $2.3M to $8.7M.

One number from the design end of the clock makes the deeper argument. Traditional pre-construction explores three or four site concepts, because a human draws each one; Marengo (YC S26) says its tooling explores more than a thousand in parallel and cuts the cycle to five or six months. The bottleneck was never engineering judgment. It was how many options one engineer can hold at once. That failure mode does not stop at the fence line.

Where the 34 days sits, drawn to scale

grid-side design-side ours

Three of these bars belong to somebody else. The orange one is Berkeley Lab's median from interconnection request to commercial operation for generation built in 2025, and it is a proxy: large-load interconnection is a separate queue that study does not measure. Use it for the order of magnitude, not the decimal.

Nothing on the teal bar is cheap because it is short. It is the last gate before a contracted megawatt starts billing, it is the one gate an operator owns outright, and it is the only one where a week of work shows up in this quarter's revenue.

Counterpoint Halving design time moves nothing if the critical path is the interconnect, and often it is. That objection is the Flux argument stated backwards: when new power is queued behind years, a watt recovered inside a connection you already hold is worth more than a watt you are waiting on.

Crucible · one engine across all five layers

Every module below answers with one recommendation: drain these nodes, migrate this job, cap power here. That is the three-or-four-concepts habit, moved indoors. Crucible runs the candidate plans against the fleet twin and hands back the trade-off frontier instead, so the operator picks the risk appetite rather than accepting ours.

Decision	Space	Traded against	Today
Maintenance window	37 nodes × 6 windows	revenue at risk · spares · tech hours	one drain list
Job placement	~10⁴ migrations	stranded HBM · fragmentation · SLA jitter	greedy heuristic
Oversubscription	ratio × class × tier	runtime penalty · SLA credits · revenue per GPU	one fixed ratio
Model residency	models × chip groups × swap windows	swap cost · traffic forecast · tail latency	pinned at deploy
Request routing	requests × engines × admission policy	capacity · fairness · tail latency	least-loaded
Power envelope	24 h × 3 tiers	energy cost · 4CP position · SLA risk	one plan
Bring-up order	rack × contract start	days to first revenue · burn-in confidence	first in, first out

concept

Not a sixth layer

Forge through Flux stay five products in a fixed order. Crucible is the search that runs over all of them, and it only works once the data plane underneath is real, which is what the land-and-expand motion builds first.

LAYER 01 Forge Hardware Engineering: NPI, custom server design, GPU validation, co-design

Nebius, on receiving its first Vera Rubin NVL72 systems, describes using them to “validate compute, networking, and orchestration together as a single system” before offering them in production (Q2 2026 letter). That is this layer, described by an operator.

Every burn-in, every validation run, every RMA becomes structured data. Forge closes the loop between what fails in the fleet and what the ODM builds next.

Burn-in first-pass yield by SKU · last 30 days

flagged: yield below 93% gate, auto-opened ECN w/ ODM

Failure fingerprints · validation + fleet, merged

Signature	30 d	Trend	Attribution
HBM ECC double-bit	214	▼ 18%	vendor lot 24-C · RMA batch open
NVLink CRC storm	96	▲ 41%	switch fw 2.4.1 · rollback queued
VRM thermal drift	61	▲ 12%	E1.S custom node · ECN filed
PCIe retimer dropout	23	▼ 6%	cable batch replaced wk 28

Use case Meta’s Llama 3 run logged a failure every ~3 h on 16,384 H100s, of which ~54% traced to GPU/HBM (arXiv:2407.21783). Forge feeds those fingerprints to the ODM in weeks; burn-in catches the 3–8% infant mortality before it ships.

LAYER 02 Fleet Datacenter Infrastructure & Fleet Management: DCIM, fleet health, node lifecycle

One health score per node, built from DCIM, BMC, and job telemetry. Failure prediction drains nodes before the job dies, not after.

Fleet health · Hall B · 512 racks (each cell = 4 racks)

health low → high predicted failure

Predictive maintenance queue

Node	Signal	Est. TTF	Action
b07-r112-n3	ECC error slope 4.2×/day	31 h	drain now
b02-r018-n7	HBM temp drift +6 °C/wk	58 h	drain at ckpt
a11-r201-n2	NVLink retrain events ×9	64 h	drain at ckpt
c04-r077-n5	PSU ripple anomaly	6 d	watch

Use case Meta logged 419 unexpected interruptions in 54 days on 16k GPUs. At posted B200 rates (~$5/GPU-hr) that cluster burns ~$82k/hr. Draining predicted failures at checkpoints instead of mid-job is worth ≈$19M/yr.

LAYER 03 Sentinel Production Engineering: reliability, incident command, AI guardrails

Every incident becomes training data. Sentinel correlates job telemetry, fabric events, and change history into a ranked root cause before the bridge call starts.

Job interruptions per 1,000 node-days · 12 weeks

interruptions / 1k node-days (Llama 3 baseline ≈ 3.8)

Auto-RCA · last 24 h

confidence 0.91

NVLink flap cascade · Hall B

Correlated w/ switch fw 2.4.1 rollout (wk 29). 14 jobs touched. Rollback queued via change mgmt.

confidence 0.84

Checkpoint stalls · storage pool 3

Metadata server GC pauses align w/ stall windows. Suggested: pin GC off-peak.

guardrail

ADLC audit: remediation agent

312 auto-actions this week · 100% within policy · 2 escalations to human, both approved.

Use case Auto-RCA cut MTTR 55% in reference design. Reliability review packs generate themselves: SEV list, blast radius, corrective actions, trend.

LAYER 04 Atlas Cloud Platforms: capacity control plane, inventory & topology, observability, auto-remediation

A live graph of every GPU: where it is, what it's wired to, who holds it, and whether it can actually be sold. The cheapest capacity on the market is the capacity already inside the fence. Fragmentation is revenue lying on the floor, and it strands memory long before it strands GPUs.

Sellable vs stranded GPUs · 8 weeks

sellable (topology-clean) stranded (fragmented / misinventoried)

Remediation recommendations

+512 GPUs

Defrag island · Hall C, pods 7–9

3 tenants movable at next checkpoint → contiguous NVL domain restored.

+288 GPUs

Inventory drift · 36 nodes “in repair”, actually healthy

BMC + burn-in evidence attached. One-click return to sellable pool.

auto

Observability rollup

Golden-signal SLOs per pod; 3 pods burning error budget. Throttle recommendation issued.

Stranded HBM · one 8 × H100 node serving Llama 3 70B

model weights runtime KV cache in use KV reserved, untouched

Why nobody sees it

Operators sell GPU-hours, but serving is bound by memory, not by arithmetic. A tenant sizes its KV cache for the longest context it might ever get, runs at a quarter of that, and every dashboard still reads 100% allocated. GPU counters cannot see inside HBM.

Agentic traffic widens the gap. Nebius reports production inference more than tripling in Q2, with a growing share agentic, where “a single task drives many model calls” and consumption scales with the complexity of the work rather than with user count. Context lengths spread further apart, so a reservation sized for the worst case wastes more of the fleet, not less.

+55% sessions

Paged KV on the four largest inference tenants

Same nodes, same tenants. Reservation follows real context length instead of the worst case.

reprice

Bill GB-hours of HBM, not GPU-hours

Charges for the input that is actually scarce, and pays the tenant to stop over-reserving.

The strand is not an HBM fact

That 352 GB bar is one instance of a claim that survives the architecture: what an operator bills for is not what runs out. Change the silicon and the scarce unit changes with it. The strand does not go away, and on the fleets with no HBM at all it gets worse, because there is nothing to page out.

Fleet	Sold as	What binds	The strand
8 × H100	GPU-hours	HBM capacity	KV reserved for a context nobody reaches
Groq LPU	tokens	on-die SRAM residency	chips pinned to a model taking no traffic
Cerebras CS-3	system-time · tokens	the wafer and its weight stream	a wafer serving a model too small to fill it
SambaNova SN40L	tokens	which tier holds which model	dead models resident in the DDR tier

Groq is the hard case, not the exception. Weights live on-die, 230 MB of SRAM per chip, so a 70B model is pinned across hundreds of chips whether traffic arrives or not, and a deterministic schedule leaves no room to oversubscribe the gap. Cerebras holds 44 GB of SRAM on the wafer and streams weights from external memory, so its strand is granularity: nobody buys a third of a wafer. SambaNova keeps 64 GB of HBM3 per socket behind 1.5 TB of DDR5, which makes it the three-tier case rather than the HBM-free one.

limit

Shape asserted, figure withheld

The 352 GB bar rests on PagedAttention's measured 60–80% waste band for static reservation. No equivalent published measurement exists for wafer-scale or SRAM-resident fleets, so these rows carry no number until one does.

Outside number Gartner, 17 Aug 2026: inference cost per agentic workflow will rise more than fivefold through 2028, even as model prices fall. The scope is the point. Not inference cost in general, but the cost of one job: routing a task to a reasoning agent instead of a chatbot costs the provider at least five times as much, before task complexity is counted. Gartner names it the Inference Paradox, better unit economics driving total cost up, and prescribes inference tiering, routing and orchestration to protect margin. Price per token falls while cost per job multiplies, and an operator watching only revenue per GPU-hour sees neither blade. That is why TOK sits on the tile above, and why routing and admission control belong in Crucible rather than in a config file.

Outside number Cast AI measured average GPU utilization at 5% across tens of thousands of Kubernetes clusters on AWS, Azure and GCP, against the 50% it calls healthy. That is not the 4% below, and the two do not add: Cast AI counts duty cycle on GPUs already allocated in enterprise clusters, while the reference model counts sellable capacity lost to fragmentation and bad inventory in a fleet already contracted out. They point the same way. Thunder Compute raised a $13M Series A on 19 Aug 2026 to sell software that reclaims the first kind. Crusoe ships MemoryAlloy, a cluster-wide KV cache with a routing gateway, against the second. Two funded companies selling shovels for the capacity this section measures is the best evidence that it is there. Reclaiming it is one job; knowing which reclamations are safe, what they cost in SLA jitter, and who gets billed after is the other.

Use case Reference model: 4% of a 48k fleet stranded at any time. Recovering it ≈ 1,900 GPUs ≈ $29–58M/yr at posted H100→B200 on-demand rates, 70% utilization. Memory strands harder than silicon. On the inference share of the fleet, the untouched KV reservation is worth 4.4 GPUs of HBM per 8-GPU node. HBM is allocated years out, so that capacity cannot be bought at any price.

LAYER 05 Flux Flexible Compute: software-defined power, power as a control input

Power is the scarcest input in the AI buildout. A GB300 NVL72 rack draws 135–150 kW, so a 512-rack hall is a ~65 MW grid asset. Flux treats it as schedulable: price and carbon signals in, checkpoint-aware power envelopes out.

Day-ahead price vs shiftable load · today

day-ahead $/MWh · shaded = recommended curtailment window

Software-defined power plan · Hall B

Window	Envelope	Mechanism	SLA impact
00–14 h	62 MW	full rate	none
14–17 h	48 MW	ckpt + GPU freq cap on preemptible tier	none
17–24 h	62 MW	catch-up burst, deferred jobs first	none

today

Projected: $5k energy avoided · 4CP position preserved

ERCOT peak avoidance is worth $40–67k/MW-yr, or $2.5–4M/yr on this hall. Zero SLA-bearing jobs touched.

Use case Emerald AI showed a 25% power cut for 3 h w/o service disruption (Nature Energy, 2026); Google runs 1 GW of DC demand response. ERCOT 4CP avoidance alone ≈ $2.5–4M/yr on a 62 MW hall, before energy arbitrage.

Yantra.