An AI factory is decided once, in a spreadsheet, and never decided again.
Before a single GPU is bought, an enterprise has to answer which workloads, how much silicon, sited where, powered when, under whose law, and structured how. That answer takes four scarce people six to fourteen weeks, lands as a spreadsheet nobody can reproduce, and is stale within ninety days — while the capital it justifies is committed for five years.
9Stages from mandate to re-adjudication. One of them has no owner in any enterprise we have met
6–14 wksTo answer the decision today, by hand, in artifacts with no provenance and no replay
99Files rendered from one confirmed blueprint — Terraform, manifests, pinned Helm, Argo app-of-apps
4,059Test cases across 130 suites — idempotence and schema validity are tested properties, not claims
THE WORKFLOW END TO END
The AI factory workflow, as it actually runs.
Nine stages carry an enterprise from a board commitment to tokens being served. Each has an owner, an artifact and an elapsed time. Read them in order and the problem is not subtle: the four stages that decide where the capital goes are done by hand, by the scarcest people in the building, in artifacts that cannot be replayed — and the stage that would catch the decision going wrong does not exist.
AI FACTORY WORKFLOW · MANDATE TO RE-ADJUDICATIONIDLE
STAGETHE QUESTION IT ANSWERSOWNED BYTODAYWITH VYUHA
Software exists for three of the nine stages. Elapsed times are the pattern we meet in enterprises, not a measured benchmark.
THE HAND-OFF AND ITS COST
Every multiplier in this category is real. Each one moves the constraint to somebody else’s stage.
Three claims the category publishes, each measured on the stage its author owns — which is the only stage they can see. That is not a criticism. The consequence is structural: improving one stage does not remove the constraint, it relocates it, and the stage it lands on belongs to a different vendor, a different budget and a different team. Claim, consequence, finding, in that order, because the first two are an argument and only the third is evidence.
HAND-OFF · WHERE THE COST LANDSIDLE
THE CLAIM
WHAT IT MOVES
WHAT WE FIND
No vendor is named. The claims belong to the category, and the pattern is the point.
Naming the vendors would turn an argument about architecture into a comparison somebody can dismiss as positioning. These are good products and VYUHA calls rather than replaces them — but a platform that owns one stage cannot see the stage its own success relocates the problem to. Nothing in this category holds all eleven at once, which is why the shortfall has no owner and nothing reports it until a customer does.
WHERE IT BREAKS
Five failures, and four of them are structural.
None of these is a tooling gap that a better spreadsheet closes. Each is a seam between two teams who hold different units, or a quantity treated as a constant that is not one.
Stages 3 and 4 do not share a unit. Sizing produces accelerators. Siting needs kilowatts, square metres and a date. That translation is done once, informally, by whoever is in the room — and every later number inherits it.
Announced capacity enters the model as supply. A campus, a megawatt figure and a year become a planning input. It is an intention gated by an interconnection queue the provider does not control, and the queues are getting longer.
Agentic workloads are sized with chat arithmetic. One user action is treated as one inference when it is a recursive tool-invocation series with a stochastic depth. The error is roughly an order of magnitude, in the direction that leaves you short.
Hard constraints are scored rather than excluded. Put residency and cost in the same weighted model and a sufficiently cheap non-compliant site wins. Being wrong about export control is a criminal matter, not a false-positive rate.
Stage 9 has no owner. The supply graph moves monthly; the decision is annual at best. Nothing re-runs it, so the gap between what was decided and what is now true widens silently for the whole life of the asset.
VYUHA is a solver with a narrative layer. The kernel computes; the language model composes and is structurally prevented from originating a quantity — enforced by a binding registry and a numeral scanner, not by asking a model nicely in a prompt.
Physics before language
Every number on an output traces to the kernel or to a dated measurement. The narrative layer may only compose what the solver produced.
Every answer is dated
A coefficient without a measurement date is unusable rather than defaulted. The refusal to size on an unmeasured coefficient is the architecture, not a gap in it.
Negative results are outputs
Infeasibility is a schema rather than an exception, and aggregated across the corpus it becomes an index of demand nobody can currently serve.
Snapshot addressability
Every solve records the supply-graph snapshot it ran against and replays bit-identically. An answer you cannot reproduce is an opinion with a timestamp.
STAGES 2–5 THE CHAIN
Stages 2 to 5, made computable.
The four stages that take a quarter by hand are one derivation: nine links from a token profile to a set of admissible placements, each constrained by the one before it. The seam between sizing and siting disappears because there is no seam — accelerators, kilowatts, racks and an energization date are consecutive links in the same solve.
DERIVATION LEDGER · 70B FP8 CLAIMS TRIAGEIDLE
Awaiting a workload.
Illustrative figures. The engine refuses to size on an unmeasured coefficient rather than defaulting one.
WHY THE ANSWERS HOLD
Four rules that are tested, not asserted.
01
Exclusions, never penalties
Hard constraints cannot be bought off.
An ITAR residency breach is not offset by a cost saving, because there is no score anywhere that could trade them. Precedence is LEGALITY > PHYSICS > DELIVERABILITY > ECONOMICS > PREFERENCE, arbitrated and never summed.
02
Unconfirmed is a value
A site fact nobody confirmed is not a fact with a default.
GPU generation, chassis density and fabric parse as UNCONFIRMED. Plan works without them and names every gap with its remedy; apply refuses. There is no force flag, because a default provisions the wrong lab and nobody finds out until week nine.
03
Idempotence is tested
The second pass must be empty.
The reconciler returns every pass and the suite asserts the second is empty. That assertion found a real defect: a priority class emitted once per tenant produced duplicate resource ids, which would have rewritten the object on every pass for ever.
04
Escalations resolve
An unresolvable gate is how teams learn to bypass gates.
Where a question is a determination for an empowered official rather than a computation, the engine reports UNEVALUABLE and clears it when the blueprint records who determined it, when, and in what scope — not via a flag that switches the check off.
STAGE 4 ADMISSIBILITY
A check that could not run did not pass.
Two-valued admissibility has to decide what to do with the third case, and either choice is wrong. Call an untested constraint a pass and you ship an unevaluated risk; call it a failure and the gate gets bypassed within a quarter. VYUHA carries the third value all the way to the output.
ADJUDICATION · CANDIDATE PLACEMENTIDLE
PRECEDENCE
AWAITING CANDIDATE
Constraint classes are arbitrated by precedence. A lower class never offsets a higher one.
STAGE 4 ENERGIZATION
Announced capacity is not rentable capacity.
A placement that cannot be energized by the date the programme needs is not a cheaper option, it is not an option. Candidates are excluded where the power block's grid-interconnection state is not confirmed by the required date, and answers come back in megawatts and months rather than only in dollars.
INTERCONNECTION GATE · REQUIRED BY Q3 2027IDLE
confirmed interconnectionannounced, not confirmeddate the programme needs
Illustrative sites. Interconnection state and queue position are curated supply-graph data.
STAGE 2 AMPLIFICATION
Enterprises are sizing 2027 agentic estates on chat-era arithmetic.
A user action against an agentic system is not one inference. It is a stochastic recursive series of tool invocations, each carrying its own prefill and decode, and the fan-out is a distribution rather than a constant. Size the estate as though a user action were a single call and the answer is wrong by roughly an order of magnitude — in the direction that leaves you short.
ONE USER ACTION · INFERENCE COUNTIDLE
SIZED AS CHAT
0
inferences per user action
One prompt, one completion. The number a capacity plan uses when nobody asked what the agent does between the question and the answer.
SIZED AS AGENTIC
0
inferences per user action
Planning turns, tool calls, each tool result re-entering context, retries on failure, and a summarisation pass. Recursive, and stochastic in depth.
—
Illustrative draw from a fan-out distribution, not a measured figure for any named system. The architectural point is that the quantity is a distribution at all — a capacity plan that carries a single number for it has already lost the information that matters.
STAGE 6 LAB FABRIC
Bare metal to day-two operations, from one declarative file.
The decision half answers where. The deployment half builds it. Ten composable, GitOps-versioned layers, each a pure function from blueprint to declared resources — so the next lab is a copy of the file rather than a new project.
RECONCILER · ONE BLUEPRINT · CONFIRMED SITE FACTSIDLE
It renders a repository, not a cluster. Argo CD converges the cluster from that repository.
Why a repository, not a cluster
A reconciler that also wrote to the cluster would be a second way for the lab to change, one that leaves no commit. The git log would stop being the complete record of intent, which is the whole reason the audit trail is worth anything.
The policy engine has a caller
An admission webhook that fails closed under strict mode, reads personnel status from the API server's own user info rather than from a pod annotation a tenant could write, and records every decision in the chained ledger before returning it.
The portal competes with its own promises
Outstanding grants are subtracted from available capacity, because observed utilisation lags every grant by however long the holder takes to start work. That is how a self-service platform approves six requests in a morning that all collide in the afternoon.
No SKU is hardcoded
Every hardware fact comes from a generation table keyed by the string the blueprint supplied, with per-entry provenance. Lookup is exact or case-insensitive, never fuzzy. A MIG request on a generation whose geometry is not recorded reports untested rather than being validated against a neighbour's profiles.
STAGE 7 THE EGRESS CALL
It does not read the prompt. It cannot, and neither can anything else.
The admission webhook approved the pod days ago. Procurement approved the endpoint last quarter. Between them sits this request, from this engineer, carrying this payload — and nothing in the stack decides it.
VYUHA decides it on three things: the workload's declared classification, the principal's asserted attributes, and the destination's recorded assessment. The API has no field for prompt text at all.
Every competitor ships a classifier that reads the payload and judges it. A classifier saying "this looks unclassified" is a probabilistic opinion standing exactly where a legal determination belongs, and being wrong about export control is a criminal matter rather than a false-positive rate.
An undeclared workload is denied here. It is not scanned, not inferred, and not waved through on a confidence score.
INPUT
SOURCE
WHY NOT THE ALTERNATIVE
declared
The workload's own classification, recorded in the blueprint
Inferring classification from content substitutes a probability for a determination somebody is accountable for
asserted
The principal's attributes, read from the API server's user info
A pod annotation is written by the tenant being governed
recorded
The destination's assessment, curated and dated
An endpoint approved last quarter is not an endpoint assessed for this payload class
STAGES 8–9 EVIDENCE
Git records what was declared. The ledger records what was attempted.
The hash-chained ledger records what was attempted, by whom, on whose authority, against what evidence. A deleted entry breaks the chain at the following entry; an edited one no longer hashes to its digest.
Tamper-evident rather than tamper-proof, and the distinction is the point: an auditor who does not trust the operator can still check it. That is a weaker claim than most vendors make and a more useful one.
Attribution says what it cannot attribute
Exclusive and partitioned consumption is exact. Time-sliced consumption shared between tenants carries a band bounded by the two extremes the telemetry cannot rule out, the band is printed on the statement, and unlabelled busy time stays in the residual instead of being spread across tenants to make the rows sum to a hundred.
What is not built
Listed in the runbooks rather than left to be discovered in a procurement call.
Nothing has been applied to real hardware.
The reference blueprint's site facts are deliberately unconfirmed, which is what makes it a useful demonstration of the gate rather than a staged success.
Newest-generation partition geometry is interpolated and cannot gate a commitment.
The demonstration telemetry is a deterministic fake, and says so.
A green cell resting on a placeholder is a number somebody signs a purchase order against. Where a figure rests on an unmeasured coefficient the arithmetic still runs, the figures are still shown, they are labelled unproven, and every finding derived from them is withdrawn — except where it holds at the best coefficient physics allows, because a shortfall the memory floor forces is geometry rather than a guess.
PORTFOLIO BOUNDARIES
Three platforms, one seam each.
The boundaries are maintained deliberately. Blurring them would weaken each position, and a platform that reimplements its neighbour acquires its neighbour's bugs without its neighbour's test suite.
VYUHA — the decision
Where a workload should run, on what silicon, under whose jurisdiction, energized when. Then the lab fabric that builds it and keeps it converged.
InferPulse — runtime economics
Owns unit economics and workload efficiency at runtime. VYUHA calls it. It does not reimplement it.
VARC — runtime governance
Owns behavioural governance of what runs. VYUHA hands off the sovereignty attestation at deployment; rule packs are a shared asset maintained once.
The corpus is the product
Software exists to accumulate the benchmark corpus, the interconnection history and the workload genome. The engine is how the corpus gets built.
INSIGHTS
Written from inside the solver.
Arguments we are actually having with infrastructure and risk teams, not thought leadership.
NEXT
Bring a workload and a deadline.
The useful first conversation is not a demonstration. Name a workload you are about to size, the date it has to be serving, and the jurisdiction it has to sit in — and we will run the chain against it and show you where it closes and where it does not.
Every other stage of the AI factory workflow has a name on it. The one that would catch the decision going wrong does not.
Write out the workflow that takes an enterprise from a board commitment to tokens being served and you get nine stages. Mandate. Demand. Sizing. Siting. Structure. Build. Admit. Account. And then the ninth, which is simply: is this still the right decision.
Eight of those have an owner. You can name the person, find their calendar, see their artifact. The ninth has none. Not a weak owner, not a contested one — none. It is on no org chart I have been shown, in no RACI, in no operating rhythm.
Why it went missing
Nothing conspired to remove it. It was never created, because the workflow was inherited from a world where the answer held.
Enterprise infrastructure decisions used to be stable for their asset life. You sized a VMware estate, you bought it, and the assumptions underneath it moved slowly enough that a three-year refresh cycle was adequate governance. Nobody needed a standing stage to ask whether the decision had gone stale, because it mostly had not.
AI infrastructure does not behave that way. The accelerator generation turns over faster than the procurement cycle. Interconnection queues lengthen. Jurisdictional rules on where models and data may sit are being written while the capital is being spent. And the workload itself changes shape underneath you, most sharply when a chat deployment becomes an agentic one and the inference count per user action moves by an order of magnitude.
So the decision now decays, and the workflow that produced it has no step that notices.
What the absence costs
The cost is invisible by construction, which is why it goes unmanaged. There is no incident, no outage, no failed audit. There is only a widening gap between what was assumed in a spreadsheet in March and what is true in the market in November, and the gap is funded for five years.
Occasionally it surfaces. A programme discovers that the site it committed to cannot be energized until two quarters after it needed to be live. An estate sized for chat traffic falls over the first week the agents go into production. A jurisdictional rule turns a cost-optimal placement into a compliance finding. Each gets treated as a surprise and a post-mortem. None gets treated as what it is: the predictable output of a workflow with no re-adjudication stage.
Why nobody has simply added it
Because at the moment, re-adjudicating means re-running stages two through five by hand. That is the same four scarce people and the same six to fourteen weeks that produced the original answer. No organisation is going to spend a quarter of specialist time per quarter to check an answer, so the honest institutional response is to not check it.
That is the actual blocker. It is not that nobody wants to know; it is that knowing is priced at the cost of the original decision, every time.
Which means the ninth stage cannot be created by assigning it to someone. It only becomes possible once the four stages underneath it are a solve against a dated snapshot rather than a bespoke piece of consulting — because then re-adjudication is running the same solve against a newer snapshot and reading the diff, and the question stops costing a quarter to ask.
Vyasa Murthy is the founder of Venture Vertex LLC. VYUHA — physics-constrained placement for AI workloads.
PLACEMENT · NOVEMBER 2026 · 6 MIN
Announced capacity is not rentable capacity
The cheapest placement in the comparison is frequently the one that cannot be energized until after the programme needed it.
Every placement comparison I have been shown in the last year leads with price per GPU-hour. It is the column everyone has, which is the only real reason it is first. It is also the column least likely to decide whether the programme succeeds, because it answers a question that only matters once a prior question has been answered affirmatively: can this capacity actually be serving by the date we need it.
The industry has learned to treat announced capacity as a planning input. A provider announces a campus, a megawatt figure and a year, and that figure enters somebody's model as supply. It is not supply. It is an intention, gated by an interconnection queue that the provider does not control, in a process whose timelines have been getting longer rather than shorter.
What the gate actually checks
The useful test is narrow and unglamorous. For each candidate placement, is the power block's grid-interconnection state confirmed by the date the programme requires, and if it is not, the candidate is excluded. Not penalised, not flagged amber, not carried forward with a caveat in a footnote. Excluded, because a placement that cannot be energized in time is not a worse option than the others, it is not an option at all.
This is the same discipline as treating hard constraints as exclusions rather than penalties, and it fails in the same way when you relax it. The moment energization becomes a weighted factor, a sufficiently attractive price can outvote it, and the model will cheerfully recommend a site that will be ready eight months after the business needed to be live.
Why answers come back in megawatts and months
Once energization is a gate rather than a factor, the shape of the answer changes. The output is no longer a ranked list of dollar figures. It is a smaller set of placements that close at all, each carrying the power envelope it needs and the date it can meet, and the dollar comparison happens inside that set.
Executives find this uncomfortable at first because it removes options. That is the point. The options it removes were never available; they were artefacts of a model that treated an announcement as a commitment.
The second-order effect
There is a consequence that only shows up once you have run the gate across many programmes. Infeasibility stops being an error and becomes data. Aggregate the placements that could not close, and across a corpus you have a measure of demand that the market cannot currently serve, broken down by the constraint that blocked it — jurisdiction, power, delivery date, or silicon that does not exist yet in the quantity required.
Nobody publishes that number because nobody is structured to collect it. A tool that treats infeasibility as an exception throws the information away at the moment it is generated. Treating it as a schema instead costs nothing and compounds.
Vyasa Murthy is the founder of Venture Vertex LLC. VYUHA — physics-constrained placement for AI workloads.
SIZING · NOVEMBER 2026 · 5 MIN
Your agentic estate is not a chat estate with more users
One user action is a recursive tool-invocation series with a stochastic depth, and capacity plans carry a single number for it.
Ask an infrastructure team how they sized their 2027 agentic estate and you will usually get a defensible-sounding chain: expected users, actions per user per day, tokens per action, tokens per second per accelerator, accelerators. Each step is arithmetic nobody can argue with. The answer is still wrong, often by close to an order of magnitude, and it is wrong at the second step.
In a chat system, a user action is one inference. Prompt in, completion out. The mapping is one to one and the arithmetic above is sound.
In an agentic system it is nothing of the sort. A single user action triggers a planning turn, then a tool call, then the tool result re-entering context, then another planning turn conditioned on that result, possibly a retry when the tool fails, possibly a sub-agent with its own series, and a summarisation pass at the end. The user sees one question and one answer. The accelerators saw somewhere between four and forty inferences, each with its own prefill against a context that grew at every step.
The distribution is the finding
The instinct at this point is to find the multiplier — to decide that agentic actions cost, say, twelve times a chat action, and carry on with the original arithmetic. That is better than nothing and it still throws away the thing that matters, which is that the fan-out is a distribution rather than a constant.
Depth varies with the query. A simple lookup resolves in two turns; an ambiguous request that requires three tools and a retry resolves in twenty. Capacity is not sized against the mean of that distribution, because the estate has to stay up during the hours when the hard queries cluster. It is sized against a tail, and you cannot reason about a tail you have replaced with an average.
Worse, the prefill cost grows with depth. Each tool result re-enters context, so the fifth inference in a series is processing a materially longer prompt than the first. The series does not cost depth multiplied by the first call; it costs more than that, and the relationship is not linear.
Why it fails in the direction it does
Sizing errors are not symmetric in their consequences. An estate sized too large wastes capital, which is visible, embarrassing and survivable. An estate sized too small fails during exactly the periods when the hard queries arrive, which is when the system is most load-bearing and most watched.
Chat-era arithmetic fails in the second direction, every time, because every simplification it makes — one inference per action, mean rather than tail, constant prefill — moves the answer the same way.
What to do instead
Size the series, not the call. Characterise the fan-out distribution for the workload class rather than assuming a multiplier, carry the growing context cost explicitly, and plan against a tail percentile that you have chosen deliberately rather than inherited by accident.
None of that is difficult arithmetic. It is just arithmetic nobody does, because the first model anyone built was a chat model and the spreadsheet was never revisited when the workload changed underneath it.
Vyasa Murthy is the founder of Venture Vertex LLC. VYUHA — physics-constrained placement for AI workloads.
ADMISSIBILITY · NOVEMBER 2026 · 5 MIN
The third value
Every admissibility check has three outcomes. Most tools carry two, and what they do with the third decides whether anyone still uses the gate a year later.
A constraint check can come back three ways. It passed. It failed. Or it could not be evaluated — the fact it depends on was never confirmed, the geometry was never recorded, or the question is a legal determination rather than a computation and no amount of compute resolves it.
Almost every tool in this space carries a boolean. Which means that at the moment the third case arises, something has to decide which of the two buckets it falls into, and both choices are wrong in ways that take months to surface.
Calling it a pass
Map unevaluable onto pass and the system becomes confidently incorrect. The report is green. The constraint was never tested. Nobody who reads the output can distinguish a check that ran and succeeded from a check that could not run at all, because by the time it reaches them the distinction has been erased.
This is the failure mode that produces the worst outcomes, because it is invisible precisely in proportion to how much the organisation trusts the tool. The more authoritative the green, the further the unevaluated risk travels.
Calling it a failure
Map it onto failure instead and the system becomes correct and unusable, which is a slower route to the same place. Teams encounter gates that block on questions the engine cannot answer and that no amount of engineering effort will clear, because the blocker is a determination awaiting a human with authority.
What happens next is predictable. Somebody adds a flag to skip the check. The flag is used for the legitimate case, then for the borderline case, then by default. Within two quarters the gate is nominally in force and functionally disabled, and nobody decided to disable it. An unresolvable gate is how teams learn to bypass gates.
Carrying it
The alternative is to make unevaluable a first-class verdict that travels all the way to the output, and to give it a route to resolution that is not a bypass.
In practice that means three things. The verdict appears on the report as itself, so a reader can see that nine checks passed and two could not be assessed. The engine names the specific missing fact and its remedy rather than reporting a general failure. And the resolution path records who made the determination, when, and in what scope — so the check clears because a question was answered by somebody accountable, not because a flag switched it off.
That last distinction is the whole thing. A gate cleared by an authority creates a record. A gate cleared by a flag creates nothing, and the next person to read the output cannot tell which happened.
Vyasa Murthy is the founder of Venture Vertex LLC. VYUHA — physics-constrained placement for AI workloads.