We built an economy that runs on AI and

now there isn't enough compute to run

that economy. A structural crisis is

emerging in global technology

infrastructure. Over the past three

years, the world economy has reorganized

itself around AI capabilities. It's now

the biggest capex project in human

history. And those capabilities depended

entirely on inference compute. That

compute is now physically constrained

and no relief is expected before 20.

Discussion documents the nature of what

is going on. explains why it differs

from previous technology supply

crunches. And we're going to analyze the

strategic implications for enterprises

and provides some actionable guidance

for leaders who have to figure out how

to navigate the next 24 months. Just

before we get into it, I'm not going to

hide the ball. These are the top things

that are popping up for me as I dig into

this issue. Number one, demand remains

exponential and uncapped. That is the

first big driver here. Enterprise AI

consumption is growing at least at 10x

annually driven by per worker usage

increases and the proliferation of

agentic systems. There is no ceiling.

Number two, supply is physically

constrained all the way through 2028 at

least. DRAM fabrication takes 3 to four

years at semiconductor capacity is fully

allocated. High bandwidth memory is sold

out. New capacity literally cannot

arrive fast enough. Number three, all of

the hyperscalers are hoarding. Google,

Microsoft, Amazon, Meta, they've all

locked up compute allocation for years

in advance. So have open AI. So is

anthropic. They are using it to power

their own products while enterprises

compete for what remains. Number four,

pricing is going to spike. It will not

rise gradually. When demand exceeds

supply this structurally and this

severely, markets do not clear very

smoothly. Trend Force projects memory

costs alone will add 40 to 60% to

inference infrastructure in the first

half of 2026. Effective inference costs

could double or triple within 18 months.

Next, traditional planning frameworks

are broken. Capex models, depreciation

schedules, and multi-year procurement

cycles all assume predictable demand and

available supply. Neither assumption

holds in the age of AI. And finally, the

window to secure capacity is closing.

Enterprises that move now can lock in

allocation before the crisis peaks. And

those that wait are going to find

themselves bidding against each other

for scraps at best or be locked out

entirely. I want to repeat this is not a

tech problem. It's being presented as

one, but that's incorrect. It's actually

an economic transformation with

consequences that will reshape

competitive dynamics across every

industry. To understand this crisis,

let's start with what is being consumed.

A knowledge worker using AI tools

aggressively, code completion, document

analysis, research assistance, meeting

summarization, consumes, I don't know,

call it a billion tokens a year if

you're really leaning in. That's the

current baseline for heavy users at AI

forward enterprises. And the ceiling is

much higher. The ceiling is 25 billion

tokens or more a year. 25x that 1

billion. And 1 billion tokens does sound

like a lot except that it isn't. A

single complex analysis task like review

these 50 docs and synthesize their key

findings can consume a half a million

tokens real easy. Day of active coding

can run into the millions of tokens. A

research session with multiple

iterations you can get as high as 10

million tokens. Consumption continues to

grow because we have not found a limit

to what we want to do with artificial

intelligence. There is no demand limit

for intelligence. Three dynamics are

driving continued acceleration of

consumption here. First, capability is

unlocking usage. As models improve, as

orchestration improves, users are

finding new applications and features

that didn't work well a year ago, like

complex reasoning, now work very well.

Every capability improvement unlocks a

lot more new demand, like nonlinear new

demand. Think about what happens when

you go from managing one agent to 10,

10x the demand. Integration is also

multiplying our touch points here. So AI

is no longer a standalone tool you open

when you need it. It is now embedded

across email clients, document editors,

development environments, CRM, etc.

Every time we integrate, we are creating

a portal for ambient continuous

consumption and agents are compounding

on top of everything. The shift from

human in the loop to agentic systems

where AI calls AI in ever more automated

loops that is that is not just a step

change. I struggle to explain how big a

change that is in consumption terms. It

is a multiple order of magnitude change.

A single agentic workflow can consume

more tokens in an hour than a human

generates in a month. Current

projections I'm just it's not that

unusual to think that a average worker

consumption model will hit 10 billion

tokens annually within the next year and

a half on average right and you may be

over a 100red billion tokens very easily

with top users. That's just

extrapolation from observed growth at AI

forward enterprises is not all that

aggressive. Some organizations are going

to be approaching those levels in

specific roles right now. Well, now

let's step back and look at what this

math means at enterprise scale. Let's

say you have a 10,000 person

organization. You're at a billion tokens

per worker. And so you're consuming 10

trillion tokens a year at a blended API

rate of two bucks per million tokens.

You're spending $20 million a year on

inference. That's expensive, but it is

manageable for a Fortune 500. At 10

billion tokens per worker, the same

organization is now going to consume a

100red trillion tokens a year. And at

the same rate, that's $200 million a

year. Suddenly, at 100 billion tokens

per year, which Agentic systems could

reach within 18 months, now your compute

bill is $2 billion.

And these calculations assume stable

pricing and available capacity, but

really neither assumption holds. It's

even more chaotic than that. The agentic

shift deserves special attention here

because it changes our consumption

models fundamentally. Human users have

natural rate limits. They type at a

certain speed. They take breaks. They

attend meetings. They go home. A human

would probably have difficulty

sustaining more than maybe 50 million

tokens a day, even working intensively.

Agents do not have such limits. An

agentic system running continuously,

monitoring, analyzing, responding,

planning. It is not far-fetched to

imagine a single worker driving that

system to consume billions of tokens per

day. And a fleet of agents working in

parallel could consume trillions. That's

not hypothetical. Enterprises are

already deploying Aentic systems for

code review, for security monitoring,

for customer service, for document

processing. And every deployment creates

sustained 247 inference demand that

dwarfs what human users generate.

that dwarfs what human users generate.

The enterprises planning for a billion

tokens a worker are planning for the

wrong curve. They need to plan for the

workers plus the agents whose workers

deploy plus the agents the enterprise

deploys centrally. This total

consumption footprint could be 10 to

100x the per human calculation. One data

point really crystallizes the insane

trajectory that we're all on. Google has

publicly disclosed that it processed 1.3

quadrillion tokens per month across its

services. That is a 130fold increase in

just over a year. Google is the world's

most sophisticated operator of AI

infrastructure with more visibility into

demand patterns than anybody else. Their

consumption curve is your leading

indicator. If the world's largest and

most capable AI operator is seeing X

annual growth, enterprise planning for

10X may well be conservative. And yet

and yet we live in a memory bottleneck.

Right? All of this assumes that we can

get the hardware to run it. But AI

inference is memory bound. The size of

the model you can run, the speed at

which you can run it, and the number of

concurrent users you can serve all

depend on memories. Specifically, high

bandwidth memory for data center

inference and DDR5 for everyone else.

And the memory market is broken. Server

DRAM prices have risen at least 50%

through 2025, likely more. Trend Force

projects they're rising another 55 to

60% quarter over quarter in Q1 2026. We

are looking at tripledigit increases in

DRAM prices and I think it goes higher

than that. The DDR5 64 gig RD dim

modules the workhorse of enterprise data

center deployment could cost twice as

much by the end of 2026 as they did in

early 2025. Counterpoint Research

projects DRAM prices overall are going

to rise, call it 50% 47% in 2026 due to

significant under supply. Look, I don't

care whether you pick the 47% number or

the 55% number. When you're talking

about prices rising this fast, this

much, this is not a typical cyclical

shortage. There are structural factors

that make this different. First, the

three big players that control 95% of

global memory production are all

reallocating away from consumer. That's

Samsung, KH Highix, and Micron. They're

all headed to enterprise segments and

specifically AI data center customers.

It's not that they're producing less,

it's that they're producing different

memory and the hyperscalers are buying

all of it. Then there's HBM

concentration. High bandwidth memory is

essential for large model inference. It

is a specialized product. SK Highix

dominates and their output is allocated

to Nvidia, AMD and hyperscalers. You

can't get HBM at any price. It's simply

not available. And then last but not

least, new capacity is not quick to come

online. It takes years. A new DRM fab

facility is going to cost you on the

order of $20 billion and it takes 3 to

four years to construct and ramp.

Decisions made today to invest will not

yield chips until about 2030. There is

no near-term supply response. Samsung's

president has stated publicly that

memory shortages will affect pricing

industrywide through 2026 and beyond.

That's the world's largest memory

manufacturer telling you on the record

they cannot meet demand. Below memory,

there's another constraint. That's the

semiconductor fab layer, and it's even

more bottlenecked than memory. TSMC

manufactures the world's most advanced

chips, including the Nvidia data center

GPUs. Their nodes, the 5nanmter, 4 and 3

nanometer, are all fully allocated.

Nvidia is their largest consumer. Apple

is second, and the hyperscalers fill out

the rest. TSMC's capacity expansion is

proceeding, but very slowly. Their

Arizona FAB won't reach full production

until 2028. New facilities in Japan and

in Germany are on similar timelines.

Intel's 18A process, demonstrated at CES

2026, represents the first credible

American alternative to TSMC. But

Intel's foundry services are unproven at

scale. Their capacity is quite limited,

and their first major customers,

including Microsoft, are going to absorb

all of their initial allocation.

Meanwhile, Samsung's foundry business

has struggled with yields on advanced

nodes, making them a less reliable

alternative for cutting edge chips. The

results is kind of predictable.

Essentially, all advanced AI chip

production runs through TSMC in Taiwan.

There is no surge capacity. There is no

alternative. This is not a short-term

fix situation either. But we're not done

with the bottlenecks. We also have a GPU

allocation crisis. Nvidia dominates AI

training and inference chips with

roughly 80% market share. Their H100 and

newer Blackwell GPUs are the standard

for data center AI. Both are sold out.

Lead times for large GPU orders exceed 6

months. The hyperscalers have locked up

allocation through multi-year purchase

agreements worth tens of billions of

dollars. Microsoft, Google, Amazon,

Meta, and Oracle have collectively

committed to hundreds of billions in

Nvidia purchases over the next several

years because of the demand I've been

talking about. And enterprise buyers are

left with whatever allocation remains,

which increasingly is not much. NVIDIA's

H200 and Blackwell GPUs, which offer

significant performance enhancements,

are even more constrained. Initial

production runs are fully allocated to

hyperscalers. Enterprise availability in

any kind of volume is very uncertain,

and you don't have great alternatives.

AMD's Instinct MI300X is competitive on

specs and available in somewhat larger

quantities, but the software ecosystem

maturity lags significantly versus

Nvidia. Intel's Gaudy accelerators have

struggled to gain market share despite

competitive pricing. Software and

ecosystem adoption remain challenges.

Custom silicone, like Google's TPU or

Amazon's Tranium, they're not available

to enterprises and they're typically

built for internal use unless you have

very special deals. None of these

alternatives changes the fundamental

picture. The GPUs that enterprises need

are controlled by companies that have

every incentive to use them internally

rather than sell them. And this is the

piece that most analyses miss. AWS,

Azure, and Google Cloud are not neutral.

They are AI product companies that

happen to sell infrastructure. They

compete directly with their enterprise

customers. Google uses its compute

allocation to power Gemini, which

competes with every enterprise AI

deployment. Microsoft uses its

allocation for co-pilot which competes

with every enterprise productivity AI

and Amazon uses its allocation for AWS

AI services competing across the board.

When compute is abundant this conflict

of interest is very manageable. The

hyperscalers can serve their own needs

and sell excess capacity and everybody

wins. When compute is scarce like now

the conflict becomes zero sum real fast.

Every GPU allocated to an enterprise

customer is a GPU not available for

Gemini, Copilot or Alexa. The

hyperscalers must choose between their

own products and their customers. This

dynamic, by the way, is true for OpenAI

and Anthropic as well. If the GPUs are

not in an OpenAI data center, OpenAI is

missing out on serving chat GPT

customers. And when in doubt, all of

these hyperscalers will choose their

products. They already have. Look at

what's happening with rate limits. API

pricing has fallen over the past two

years, but rate limits have tightened.

Enterprise customers report increasing

difficulty getting allocation

commitments for high volume deployments.

And to be honest, hyperscalers are not

being the villains here. They're being

rational. Their AI products are

strategic priorities with internal

champions. And selling capacity to

enterprises is just a business. It's not

the business their leadership is

necessarily measured on. In the race to

AGI, enterprise CTOs need to internalize

that the cloud providers are not going

to be reliable partners in this crisis.

They are competitors who control the

scarce resource that you need. Now, of

course, in a smoothly functioning

market, prices will rise gradually as

supply tightens, demand will moderate,

and equilibrium is restored. This is not

the world we live in. Because supply

cannot respond to demand and demand

cannot be deferred, prices are going to

spike. Buyers will bid against each

other. They're willing to pay premiums.

Sellers seeing the desperation will

raise prices for extraction, not for

equilibrium. We've already seen this

before. DRAM prices spiked 300% during

the 2016 shortage. GPU prices doubled

during the crypto mining boom. Memory

prices are notoriously volatile because

the supply side is so inelastic. You

just can't produce chips fast. And the

current situation has all the

ingredients for a severe spike. Supply

is inelastic. No new fabs coming online

for a couple years. Demand is inelastic.

Enterprises are committed to AI and

they're not changing. Information is

asymmetric. The hyperscalers know how

constrained supply is. Enterprises by

and large do not. And coordination is

possible with only three major memory

suppliers and one GPU supplier. Tacic

coordination it can happen and no

antitrust violation is required to do

that. Now the impact of a spike in

inference cost is going to depend on

your business model. AI native startups

are extremely exposed. Companies like

Notion have publicly disclosed that AI

costs now consume 10 percentage points

of what was previously a 90% gross

margin business. If AI is margin diluted

at current pricing and inference costs

double, many AI native business models

are going to become unviable. Enterprise

software companies building AI features

face similar pressure. AI is a

competitive requirement, but every AI

feature erodess margin. Companies are

going to have to choose between

competitive necessity and financial

sustainability. Enterprises using AI

internally have somewhat more

flexibility. If AI is creating

disproportionate value, the cost

increase can be justified, but I would

expect budget scrutiny to intens

intensify, but I would expect budget

scrutiny to intensify over the next two

years. AI projects that were approved at

one cost level may well be canceled at

twice the cost. Now, the hyperscalers

themselves, ironically, are somewhat

insulated. They own the infrastructure

that's becoming scarce, but even they

will face constraints. Google,

Microsoft, and Amazon are all warning

investors about rising AI infra costs.

The companies most at risk are those in

the middle. They're too dependent on AI

to abandon it. They're not large enough

to secure dedicated allocation, and they

are competing in markets where pass

through cost increases are very

difficult to sustain. So, let's cut to

the chase. How do you plan for this if

you're in the enterprise? That's the

billiondoll question. traditional

planning fails. So enterprise IT

planning evolved for a fundamentally

different era and is just not ready. The

traditional model was to assess

requirements to procure infrastructure

to depreciate it over 3 to 5 years and

all of that assume predictable demand

technology that was stable and supply

that was available. None of that is true

anymore. Demand is unpredictable and

exponentially scaling. Technology itself

is unstable because model architectures

and hardware capabilities are changing

rapidly and supply is extremely

constrained. As we've discussed, CTO's

who apply traditional planning

frameworks in that environment are going

to set themselves up to systematically

make bad decisions. They're going to

overcommit to long-term purchases that

become stranded assets. I've seen it

already. They're going to underinvest in

flexibility and optionality and they're

going to assume supply availability that

isn't really there. Let's consider a

real example to make this sort of

tangible and concrete. Let's suppose we

have an enterprise that purchases a

thousand AI workstations with NPU

capabilities at 5 grand each. That's a

$5 million capital investment. Finance

sets a 4-year depreciation schedule. It

expects to extract 1.25 million a year

in value. By year two, those same

workstations cannot handle the workload

because per worker consumption has gone

up 10x. The NPUs that were adequate for

code completion and document

summarization cannot sustain agentic

workflows consuming billions of tokens.

The machines aren't broken, they're just

obsolete. What does the enterprise do at

this point? Option A is continue using

inadequate hardware. Workers get

constrained. Productivity growth lags.

Competitors with better infra ahead. The

savings from extending depreciation cost

you much more than the lost

productivity. Option B, purchase new

hardware. The enterprise takes a right

down on the assets. The $5 million

investment yields maybe 2 million in

value and finance is unhappy. Option C,

lease instead of buy. The enterprise

pays a premium because lessers aren't

charities, but it transfers depreciation

risk. Option C may be the correct

answer, but it also passes the buck and

you have to be able to find a way to

lease tech at scale that actually works.

And so as much as it is probably the

ideal solution from a balance sheet

perspective, I have yet to see an

enterprise successfully execute a

large-scale lease. What I have seen is

large-scale commitments to cloud as a

way of deferring cost incurred at the

enterprise level. And I do wonder if one

of the ways forward here is that the

workstations we're going to be using at

the enterprise will remain pretty dumb,

but we will buy that scarce cloud

capacity, which is exactly what the

hyperscalers would want because cloud

providers will offer substantial

discounts for committed user agreements.

If you're doing a multi-year commit to

cloud, you can reduce effective pricing

by 30 to 50% compared to an ondemand

rate. at 10x annual demand growth. The

problem is that those multi-year commits

end up being traps because in scenario

one, if you're buying cloud, you can

undercommit. You estimate 10 trillion

tokens a year for the business. Your

actual consumption is 30. You pay on

demand rates for your overage and you're

in real trouble on a budget perspective.

Scenario two, you get overaggressive and

you overcommit. You estimate 30 trillion

tokens. Your actual consumption is 15.

Well, now you paid for 30 trillion

regardless and you have half your spend

as waste. Scenario three, you commit

accurately and this requires very

carefully predicting your AI

consumption, your AI capability

improvements, your efficiency gains,

etc. And the probability of accurate

prediction across the dynamic

environment we're in is in practice

zero. Option three is what people wish

was there and it's not really there. A

lot of enterprises are looking at this

and they are choosing option one.

They're doing a committed use agreement

and they're treating that as the floor

and they're going with overages as

capacity grows because they can't

predict it. And that lines up with sort

of some of the the and that lines up

with the strategic levers that I've seen

sharp CTOs actually use in this kind of

environment. Principle number one, sharp

CTO are securing capacity before they

need it. The single highest impact

action an enterprise can take is

securing inference now before the crisis

peaks. That doesn't mean you're

necessarily signing a gigantic committed

use agreement. It means obtaining some

contractual guarantees of throughput

with some degree of SLAs for

availability. And so the conversation

with vendors should shift from what is

your price per million tokens to can you

contractually guarantee us x billion per

day sustained with 99.9% availability.

If the vendor can't deliver the volume,

their pricing is often irrelevant.

Principle number two, sharp CTOs are

building a routing layer. The most

durable competitive advantage in this

environment is the intelligence layer

that decides where workloads run. A

sophisticated routing system is going to

optimize for cost. It's going to manage

capacity. It's going to preserve your

optionality by abstracting the

underlying infrastructure. So switching

providers is trivial and it's going to

enable negotiating leverage. But

building this layer, it's not trivial.

You have to have the right architecture

for a routing layer. You have to be able

to evaluate models intelligently on the

fly. You have to have great

observability and you're going to have

to hire a team to sustain this. So this

is something that is so important that

if you are operating at enterprise

scale, you cannot really outsource this

capability. This is something that you

have to have the secret sauce that

connects the usage you have internally

with a smart router that optimizes for

cost. The routing layer is how you

maintain your business independence.

Principle number three, sharp CTOs treat

hardware like a consumable. Any hardware

purchased for AI workloads should be

mentally depreciated within 2 years

regardless of how accounting is going to

treat it. So for workstations and edge

devices, yeah, you should lease where

you can. It's often hard to do. And if

you're purchasing, I would use an

accelerated depreciation schedule

because I think that reflects reality.

You need to plan for refresh cycles that

coincide with hardware generations

because every 18 to 24 months there's

going to be a new GPU architecture that

arrives with a really significant

capability improvement you're going to

want. Principle number four, sharp CTOs

invest in efficiency in a supply

constrained environment. Efficiency is a

competitive advantage. Every token you

don't consume is capacity you can

allocate to additional workloads. And so

an enterprise that can accomplish the

same task with 50% fewer tokens has

twice the effective capacity. This is

why Deepseek's work on engram is so

interesting because they were able to

effectively dramatically reduce token

usage at inference for factual lookups.

But it's not just use the smallest model

capable for each task. Well-designed

prompts can result in much lower token

usage. Caching can result in much lower

token usage. Retrieval augmentation can

use embedding based retrieval that is

orders of magnitude cheaper than raw

inference. This is a little bit of what

DeepS was getting into. Quantization can

enable a smaller model to match larger

model performance on very specific

tasks. These kinds of efficiency

investments have traditionally been

lower priority than capability

investments, but in a constrained

environment, they're going to become

critical. The enterprises that can

unlock Xefficiently effectively have

given themselves 10x more capacity. The

global inference crisis if we step back

is not a prediction. I'm not predicting

something here. I am simply observing

current conditions and what is right

around the corner based on existing

price raises we're seeing. The demand

curve here is exponential. The supply

curve is flat and the gap is just going

to widen for the next couple of years.

The playbook is pretty clear when you

actually see the situation for what it

is. You have to secure your capacity

now. You have to build a routing layer

that enables you to allocate your choice

of model where you want it. You have to

treat hardware more as a consumable.

It's a big change for IT departments.

You have to invest in efficiency as if

it's your competitive advantage. And

you're gonna have to think about how you

can diversify across your entire stack

as much as you can so that you reduce

your dependence on any single player in

the ecosystem. The enterprises that act

on this playbook will be positioned to

operate through the crisis and compete

effectively when supply starts to even

out on the other side. Those that don't

are going to find themselves capacity

constrained and cost pressured and

they're going to be falling behind in

the biggest technology race in history.

But this is, remember, this is not a

technology problem. It is an economic

transformation problem. And it's going

to separate winners and losers based on

these kinds of decisions that I'm

outlining here that CTO's will make in

the next 6 months. The window is open

for action, but it's not going to stay

open long given where the prices are

going. The moment to move and secure

your capacity is now. Best of luck.