Executives in a dark boardroom, a projected spreadsheet grid on the wall

Intelligence
Debt.

Every enterprise is running less intelligence than it could. The gap is a liability. It accrues interest, and it appears on no balance sheet.

The conclusion first. We call the gap Intelligence Debt: the difference between the capability an enterprise could deploy with today's models and the capability it actually deploys, plus the cost of closing it. Several things accrue it. The largest, and the subject of this article, is model fragmentation: a hundred rational decisions that become a hundred models, each selected, secured, validated, hosted and upgraded separately, forever. In our five-year model the fragmented estate costs about 2.5× as much and still runs two generations behind. The fix is one organization-level AI strategy: a model strategy, a platform strategy, and deliberately limited business-unit strategies. The pages that follow show how the debt accrues, what it costs, and how to retire it.

Bud EcosystemOctober 202612 minute read

For CIOs, CDOs, CROs and the people who write the RFPs

Intelligence Debt · What we keep seeing

The conclusion, then the evidence

One governed layer, or interest on every model, forever.

What it is

The gap between the intelligence an enterprise could run and the intelligence it does, plus the cost of closing it. It accrues interest.

How it accrues

Model fragmentation, above all. Each business unit funds its own use cases and picks its own model; nobody sees the aggregate. 40–100 models.

What it costs

Six fixed charges per model, due at every version, plus a capability gap that widens daily. Five-year TCO: $171.6M fragmented, $68.2M governed.

What to do

One organization-level AI strategy: a model strategy (3 core roles, 6–10 specialists), a platform strategy (one gateway), limited BU strategies.

Where this comes from: the RFP desk. For eighteen months the same document has reached us from banks, insurers, health systems and ministries. One table in the RFP lists eighty to a hundred and twenty use cases. The next table asks bidders to name a model against each one.

Each row carries its own model, owner and budget line. The spreadsheet looks like diligence. But really it's a plan to run a hundred models, and the RFP never asks the two questions that would expose the cost: how is each model retired, and who pays to keep it current?

The estates are already there. Across 20,000 organizations on one data platform, 78% run two or more model families; 37% of enterprise CIOs run five or more. The model count is not going down. What is still open is how those models are organized.

80–120

use cases in a typical enterprise or government GenAI RFP we answer

1 : 1

use case to model, in roughly seven of ten requirement tables

~0

RFPs that ask how a model is retired or kept current

Composite of the RFPs Bud has answered in the last 18 months. Rounded.

From a typical RFP · Proposed model per use caseExcerpt · composite

Use case	Proposed model	Owner · budget line
Claims triage and routing	Model A	Claims ops · BU 1
KYC document extraction	Model B	Compliance · BU 2
Contact-center summarization	Model C	CX · BU 3
Internal code assistant	Model D	Engineering · IT
Policy and procedure search	Model E	Legal · BU 4
Loan pre-screening agent	Model F	Lending · BU 5
… 94 more rows	… 94 more models

Definition

Intelligence Debt is the gap between the capability an enterprise could deliver with the best available models and the capability it delivers with the models it runs, plus the cost of closing that gap. It is paid in money and in capability.

Several things accrue it: fragmented model estates, absent evaluation, data that is not AI-ready, infrastructure lock-in. This article covers the largest, model fragmentation. Later articles take up the rest.

Bud Ecosystem · Intelligence DebtWhat we keep seeing02

An office tower at night, every floor lit differently

Intelligence Debt · How it accrues

How the debt accrues

Every floor funds its own AI.

Nobody decides to run a hundred models. It happens one rational decision at a time, on different floors of the same building, each defensible on its own and none visible to the others. The business unit funding the fourth claims model has never heard of the first three, and the vendor paid per use case has no reason to mention them.

Security, procurement and model risk review each model on its own, because that is how the process is written. The use case ships with prompts tuned to that model, which makes leaving it expensive. Then the next floor starts. The aggregate is nobody's job until it is everybody's problem, and the aggregate is where the debt accrues.

1

A business unit funds a use case

Its own budget, its own timeline, its own vendor conversations. Central IT hears about it at procurement, if then.

2

A bake-off picks the quarter's winner

The model that topped the benchmark that month. Reuse of what another floor already validated is never on the scorecard.

3

The vendor is paid per use case

One contract, one endpoint, one credential set. Suggesting a shared model would shrink the deal.

4

Every review happens per model

Security, legal, licensing, model risk and procurement each clear the model in isolation. No review sees the estate.

5

The use case ships, pinned

Prompts, retrieval and guardrails are tuned to that model's quirks. Migration becomes a project nobody has funded.

×100

Repeat, floor by floor

Forty, sixty, a hundred models. Each one selected, secured, validated, hosted and upgraded separately. Forever.

Two properties make this worse than it looks. It is invisible: each use case's budget shows only its model's bill, while the duplicated validation, security and platform work lands in shared cost centers. Gartner warns that organizations misjudge GenAI costs by 500% to 1,000% for exactly this reason. And it compounds: agents are coming, 150,000 per Fortune 500 enterprise by 2028 on Gartner's estimate, and agents need one consistent model layer. A hundred siloed integrations cannot host them.

40–100

models for a
100-use-case estate

Bud Ecosystem · Intelligence DebtHow it accrues03

Intelligence Debt · The bill, itemized

The bill, itemized · Paid in money

Six charges, levied per model, due again at every version.

This page is the money half of the debt. A model is not a line item. It is a liability with a fixed cost that has nothing to do with how many tokens it serves. Every model in a regulated enterprise must be selected, security-reviewed, validated, cleared by legal, integrated, guarded, evaluated, red-teamed, monitored and hosted, then put through most of that again at every vendor version. In a one-model-per-use-case estate that cost is paid once per use case. In a portfolio it is paid once per role. We call the six charges below the Model Fragmentation Tax. They are the fixed term; the next page is the other currency, the one that never appears on an invoice.

×N

every charge below is paid per model in the estate

×1

per role, in a governed portfolio

Statement of charges · Assessed per modelSix charges, one estate

01 Hardware underutilizationCapacity is reserved per use case and sized for its peak. Cast AI measured 5% average GPU utilization across 23,000 clusters. Inference is now 55% of AI infrastructure spend, so every idle deployment is a recurring bill. 5%average GPU utilization, unoptimized clusters (Cast AI)

02 No model reusePrompts, retrieval tuning, evaluation sets, fine-tuning data and feedback are trapped inside each integration. The fifth claims use case starts from zero. Nothing compounds. 0shared assets across one-model-per-use-case integrations

03 Depreciating assetsA model is pinned at admission quality while the frontier moves monthly. Migrating is expensive because the prompts were tuned to that model; 66% of teams never switch. The estate runs on last year's intelligence by default. 1 yr / dayhow fast a pinned model ages against the frontier (page 05)

04 Security and procurement, N timesEach model is another endpoint, credential set, data path, vendor contract, concentration-risk entry and exit plan. Prompt-injection defenses, PII handling and output checks are rebuilt per integration. 97% of AI-breached organizations lacked basic AI access controls. +$670Kbreach premium when shadow AI is involved (IBM 2025)

05 No quality gatesFew use cases ship with a maintained evaluation set, so there is no objective basis to migrate, route or upgrade, only opinions. A hundred models consume the validation capacity before most use cases reach it. 41%of financial institutions say validation already blocks GenAI

06 Hard to updateEvery vendor version is a mini-migration: re-tune, re-validate, re-test, re-release, per model. CIOs call switching "a task that can take a lot of engineering time." Multiply by the inventory and upgrades stop. ×Nevery upgrade, repeated per model in the estate

Total · per model, again at every versionPaid in money and time · sources on page 10

Bud Ecosystem · Intelligence DebtThe bill, itemized04

An old accounting ledger open on a dark desk

Intelligence Debt · The capability gap

The bill, continued · Paid in capability

The estate ages a year every day.

Technical debt is the cost of code you chose not to fix. The capability half of Intelligence Debt is the cost of models you chose not to refresh. A hundred-model estate accrues it a hundred times over.

Mechanism

A pinned model keeps admission-day capability while the frontier moves. Every deferred refresh makes the next migration dearer, and every new use case built on a stale model inherits the staleness. Fragmentation multiplies the exposure: a hundred pinned models is a hundred separate refresh bills, none of them funded.

The rate is what makes it expensive. Frontier models double the length of task they can complete roughly every seven months; the price of a given capability has fallen 280× since 2022. Measured against the frontier, a pinned model ages about a year every day it is left unrefreshed. The use case still works. It just works like 2025.

Figure 1. The pinned model keeps admission-day capability while the best available model keeps moving. The shaded gap is Intelligence Debt. Illustrative curve; frontier pace from METR's 7-month doubling of task horizon.

7 mo

frontier task-horizon doubling time (METR)

280×

price decline per unit of capability since 2022 (Stanford AI Index)

66%

of teams stayed on their provider instead of switching (Menlo Ventures)

Principal

The refresh work not done: re-tuning, re-validating and re-releasing each use case on a current model.

Interest

The widening capability gap, plus every new use case built on the stale model, which now has to migrate too.

Default

Migrations deferred, the P&L fine, until a competitor's claims agent closes files in a quarter of the time. The use case is rebuilt from scratch, on yet another model.

Bud Ecosystem · Intelligence DebtThe capability gap05

Intelligence Debt · Total cost of ownership

Total cost of ownership · A worked projection

$87.4M or $38.1M over three years, for the same use cases.

The governed portfolio costs 60% less over five years, and the gap widens every year. The model is synthetic and deliberately mid-range: a regulated enterprise growing from 60 to 140 GenAI use cases, the fragmented estate admitting roughly one model per 1.4 use cases, the portfolio holding 3 core roles and a specialist tier that settles at 13. Replace the assumptions with your own and the shape survives, because fixed cost in a fragmented estate scales with the model count, and the model count scales with the use-case count. A portfolio breaks that link.

$171.6M

fragmented, five-year cumulative

$68.2M

governed portfolio, five-year cumulative

−60%

five-year difference; −56% over three

FragmentedAdmissionRun and dedicated hostingRefreshTokens

PortfolioAdmissionRun, pooledPlatformTokensEvaluation sets

Figure 2. Annual cost, USD millions, both estates serving the same use cases. The fragmented estate's bar is mostly fixed cost per model; the portfolio's is mostly platform and tokens, which is what a chargeback can actually meter.

Assumptions · synthetic, mid-range, USD

Cost line	Fragmented · one model per use case	Portfolio · roles behind one gateway
Admission, per new model	$300K once, selection through integration	$450K once, validated deep rather than wide
Run, per model-year	$220K, dedicated hosting at low utilization	$160K, pooled capacity at 60%+ utilization
Refresh, per model-year from Y2	$80K, migration and re-validation per model	Inherited behind the router
Platform, per year	None shared; rebuilt inside each integration	$4.0–4.4M, gateway to catalog, team of 8–15
Tokens, per use case-year	$90K unrouted, largest model by default	$50K routed, plus $15K for an evaluation set

Model counts: fragmented 42 → 98, portfolio 9 → 13, for 60 → 140 use cases. Refresh is modeled as paid in full here; page 07 shows what happens when it is not.

Bud Ecosystem · Intelligence DebtTotal cost of ownership06

Intelligence Debt · Below the line

The projection, continued

The estate that looks cheaper on the P&L is the one accruing the debt.

Two things the annual bar chart hides. The portfolio gets cheaper per use case as it grows; the fragmented estate does not. And the fragmented estate that reports the lowest cost is the one that has stopped paying its refresh bill, which does not make the bill go away.

Figure 3. Cost per use case per year. The portfolio falls from $223K to $111K because the platform is paid once. The fragmented estate barely moves, $454K to $322K, because every new use case brings its own fixed cost.

Finance rarely sees Figure 3 because the denominator is missing: nobody meters cost per use case in a fragmented estate. Each BU sees its own model's bill, which looks reasonable. The CFO sees GenAI "over budget," which 73% of FinOps practitioners now report, and no lever to pull.

Figure 4. The fragmented estate when refresh is deferred, which is what 66% of teams do. Above the line, what the P&L shows. Below it, the deferred-refresh component of Intelligence Debt: $20.2M of unpaid migrations by year five, on models two generations behind.

This is the quiet version of the debt. Nothing breaks. The claims agent still triages. It just does it the way the best model of two years ago did, while the enterprises on page 08 route the same work to this quarter's model without the use case noticing.

−60%

five-year cost, portfolio
versus fragmented, refresh paid

$20.2M

Intelligence Debt accrued
when refresh is deferred

The business case does not rest on routing savings. Those are sometimes overstated (one independent test measured 4.5% to 14.2% against a 60% vendor headline), so the token assumption here is conservative. Governance, security, validation capacity and time to value justify the portfolio at zero routing savings. Routing is the dividend, not the premium.

Bud Ecosystem · Intelligence DebtBelow the line07

A quiet control room, two operators facing one wall of screens

Intelligence Debt · The strategy

The strategy that retires it

One strategy for the organization, not one per floor.

Organization-level AI strategy= Model strategy+ Platform strategy+ Limited BU strategies

The fix is not a better model. It is a different unit of decision. Stop choosing models for use cases and start filling roles for the enterprise. A role is permanent: frontier reasoning, general workhorse, small and fast. The model behind it is whatever is best this quarter, and it changes behind the router without any use case noticing. Three strategies, three owners, and one authority that can say no to the hundred-and-first model.

Model strategy

Owned by a portfolio board

Roles, not models. Three core roles carry about 80% of traffic.
A specialist tier of six to ten: code, embeddings, documents and vision, speech, narrow validated models, fine-tunes as adapters.
An admission test: a new model must prove a gap no core model closes.
Validate deep, not wide; retire on a standing agenda. A slot is freed before a new model takes one.

Platform strategy

Owned by a platform team of 8–15

One gateway in front of 100% of traffic, in before anything else changes.
Routing by data class and residency first, cost and latency second; guardrails and PII policy enforced once.
Metering per use case, so cost per outcome exists for the first time.
Evaluation as the control; pooled sovereign capacity; failover between two models per role.

BU strategy, limited

Owned by use-case teams

Use cases and outcomes. The business owns what the AI is for and whether it worked.
Prompts, retrieval data and a maintained evaluation set per use case.
Consumes the portfolio through the gateway; a new use case goes live in weeks, inheriting approved models and controls.
Does not select, host, validate or contract for models.
Does not issue its own AI RFP.

9–13

models in production, down from 20 to 100

80%+

of traffic on the three core roles

100%

of GenAI traffic through one gateway, in one log

18 mo

from sprawl to a governed portfolio, four gates, no pause in delivery

Bud Ecosystem · Intelligence DebtThe strategy08

Intelligence Debt · What is settled

What is settled, and what is not

Many models is settled. A hundred integrations is a choice, and the leaders have made the other one.

The institutions furthest ahead answered the "behind what" question the same way: one governed platform, a rotating bench of three to six models, use cases consumed through the platform and never owning a model.

JPMorganChase serves 450+ production use cases through one model-agnostic layer. Goldman Sachs routes each request by task inside the perimeter. Allianz runs one front door for 60,000 users. None of them chose one model. None chose one model per use case.

The regulators are, in effect, asking for the same thing. None asks for fewer models; all ask for things that get harder with every model added: inventories, concentration management, per-system risk files, tested exit plans.

What remains is the spreadsheet, and the RFP that keeps asking for a model in every row. Change the question from "which model for this use case" to "behind what," and the hundred-model plan does not survive the first review. The test on the right is what that review asks.

The admission test

A use case earns a dedicated model only when all four hold:

Its evaluation set shows a material, documented gap against every core model after prompting, retrieval and adapters have been tried.

The gap comes from model class or modality, not configuration.

A regulator, contract or data class requires it, or volume makes a dedicated small model cheaper at measured utilization.

The board can name an owner, a successor plan and the slot, and the specialist tier stays at ten or fewer.

What to do on Monday

Count the models. Every endpoint, adapter and API key, including the ones nobody approved. Most enterprises are surprised by the number.

Put one gateway in front of all of it before changing anything else. Visibility first, decisions second.

Freeze new model admissions until three core roles are named and filled.

Give every production use case an evaluation set. Without it there is no migration, routing or upgrade, only opinions.

Name the portfolio board with the one authority that matters: to say no to the hundred-and-first model.

Rewrite the model table in the next RFP. Ask for roles, routing, evaluation and exit, not a model per row.

The spreadsheet with a model in every row is not a strategy. An ecosystem with roles, routing and governance is. Build the second and the debt stops accruing.

Bud Ecosystem · Intelligence DebtWhat is settled09

Intelligence Debt · How Bud helps

How Bud helps · The one page we promised to keep to one page

Three models, not a hundred, is a decision. Bud is how it is kept.

Everything on pages 08 and 09 can be run by hand. Bud Novaria, the Enterprise AI Operating System, is the organization-level strategy built in: gateway, portfolio control plane, pooled sovereign inference, governance and the builder for use-case teams, on one substrate, inside your own environment.

It deploys on the hardware you already own, CPUs, GPUs, HPUs or NPUs, with zero external dependencies and 1.5–3× better inference performance without hardware lock-in. Your data never leaves. And because every layer shares one substrate, the per-model boundaries the debt accrues at are not managed. They do not exist.

Strategy layer	Bud product	What it does about the debt
Platform strategy · the gateway	Bud AI Gateway	One front door for 100% of traffic: routing by policy, data class, cost and latency; guardrails once; metering per use case; one audit trail; failover between roles.
Model strategy · the control plane	Bud AI Foundry	Catalog, admission, retirement and staged promotion. Shadow-test a successor on live traffic, promote it behind the router, and every use case inherits the upgrade.
Platform strategy · pooled inference	Bud Runtime	Universal inference across CPUs, GPUs, HPUs and NPUs with multi-adapter serving; fine-tunes become adapters on a shared base and idle dedicated capacity ends. Up to 56% token savings.
Model strategy · the open-weight bench	Bud Models	Open LLM, multimodal, diffusion and code models for the workhorse, small-fast and specialist roles, hosted sovereign, with no per-use-case vendor contract.
Governance · the evidence	Bud Sentinel	Zero-trust model governance and the evidence the portfolio board, model risk and the supervisor need: one inventory, one policy, one log.
BU strategy · the paved road	Bud Studio	Where use-case teams build agents, tools, prompts and evaluation sets against the governed portfolio, in natural language, never selecting or hosting a model.

1.5–3×

Inference performance, no hardware lock-in

0

External dependencies. Your data never leaves

30 days

To a sovereign AI cloud on your hardware

Put your spreadsheet on it.

Send us the use-case list from your RFP and we will return it as a role map: which three core models would carry it, which specialists it needs, and the five-year bill both ways, on your numbers.

Map my use cases →

Sources: Databricks State of AI Agents 2026 (78% run 2+ model families, 20,000 orgs) · a16z enterprise CIO survey, Jun 2025 (37% run 5+) · Gartner (500–1,000% GenAI cost misjudgment; 150,000 agents per Fortune 500 enterprise by 2028) · Cast AI 2026 (5% GPU utilization, 23,000 clusters) · ClearML · FinOps Foundation 2026 (73% over budget) · IBM Cost of a Data Breach 2025 ($670K shadow-AI premium; 97% lacked AI access controls) · McKinsey and IACPM (41% of financial institutions, validation blocks GenAI) · Menlo Ventures 2025 (66% upgraded within provider; 11% switched) · METR 2025 (task-horizon doubling ≈7 months) · Stanford AI Index 2025 (280× price decline) · JPMorganChase, Goldman Sachs, Allianz, US GSA public statements and reporting · Bud RFP desk, composite observations across enterprise and government RFPs, 2025–2026 · TCO projection: synthetic, Bud analysis.

IntelligenceDebt.