InflectAI, Inc.

Field Note

Aftermath: What Happens After the AI Bubble Pops

The AI bubble breaks when the capital stack can no longer hide the gap between falling token prices and the infrastructure built to serve them.

  • Entity: InflectAI, Inc.
  • Published: June 30, 2026
  • Status: current

I bought Exodus Communications as a second lieutenant in the Army and lost a few thousand dollars.

The money mattered. I was young, underpaid, and reading about the internet between Army tasks. Exodus sounded like the disciplined bet. Data centers. Enterprise customers. Backup power. Servers. Network connections. A company selling the picks and shovels while everyone else chased sock puppets and banner ads.

Then Exodus filed for Chapter 11 in September 2001.

It had 44 internet data centers across four continents. It listed about $5.98 billion in assets and $4.44 billion in debt. It had been public since 1998 and had never made a profit. Wired had described it a few months earlier as the brick-and-mortar survivor of the dot-com wreckage: fortress buildings, mission-critical servers, serious customers, and enough physical infrastructure to feel safe.

The stock had traded as high as $86.

By the time the company was fighting for survival, it was a little over $2.

That was my expensive introduction to capital-intensive bubbles. You can be right about the technology and still be dead wrong about who gets paid for building it.

The internet kept going. Exodus shareholders did not.

AI is walking into the same test.

Parts I and II were about the subsidy.

Part I argued that frontier-token prices are adoption prices. The public rate card leaves the full cost stack unanswered. Part II argued that falling token prices move the pain. Customers get cheaper AI. Labs get more usage. The owner of the GPU fleet gets a machine whose earning power can reset faster than the depreciation schedule.

Part III is the bill coming due.

The AI bubble pops when real demand is financed with contracts and assets that need a smoother world than AI pricing will give them. Token prices have to fall for adoption to continue. Utilization has to rise fast enough to offset the price cuts. GPUs have to hold premium earning power long enough to justify the buildout. Frontier labs have to pay compute commitments signed when capacity was scarce.

The whole trade depends on all four working at once.

Exodus belongs here because it separates two ideas that investors love to confuse: technology adoption and capital recovery. Data Center Knowledge later noted that Exodus left behind high-end data centers that were bought or leased by companies including Savvis, Google, DuPont Fabros, Cable & Wireless, and others.

The infrastructure survived.

The equity was wiped out.

Capital-Intensive Bubbles

AI is a capital-intensive bubble.

The comparisons that teach the right lesson are railways, telecom fiber, shale, and data centers. Each one had a real technology underneath it. Each one had real demand. Each one also had a period where capital rushed into physical capacity faster than the cash flows could support.

People keep missing the financing layer. The bubble is a claim about the price paid to build the physical layer.

Railway tracks have a cost basis. Fiber has a cost basis. Shale wells have a cost basis. GPUs have a cost basis.

The next owner can make the same asset work because the first owner overpaid.

The aftermath can look strange from the outside. Customers get better service. The infrastructure stays in place. The early shareholders get buried.

The late-1990s telecom bubble is the obvious historical teacher. After the 1996 Telecommunications Act, capital poured into networks, switches, fiber, hosting, and internet infrastructure. The economy needed more bandwidth. The internet became the main commercial network of the world. The buildout outran the cash flows. WorldCom collapsed. Global Crossing filed for bankruptcy. Exodus filed for bankruptcy. A lot of fiber and data-center capacity survived the wreckage and later became valuable to someone else at a lower cost basis.

Lower cost basis is the phrase to hold onto.

The next owner can make the economics work because the first owner absorbed the loss.

AI infrastructure has the same risk. The world may need vastly more inference. Enterprises may put AI into every workflow. Agents may become normal software infrastructure. None of that tells you whether the companies buying GPUs at peak prices, signing power commitments at peak demand, and building data centers into constrained grids will earn an acceptable return on those assets.

The Cash Flow Statement Talks First

Accounting spreads the cost over time. Free cash flow takes the hit now.

Amazon is already showing the pressure. In Q1 2026, Amazon said trailing twelve-month operating cash flow rose to $148.5 billion. Free cash flow fell from $25.9 billion to $1.2 billion, driven mainly by a $59.3 billion year-over-year increase in property and equipment purchases tied primarily to AI investment.

The bubble is already moving through the cash flow statement.

CoreWeave shows the contract version. At the end of 2025, it reported $66.8 billion of revenue backlog, up from $15 billion. Backlog sounds like safety. The capacity is spoken for. The cost side is already moving. Depreciation and amortization rose from $843 million in 2024 to $2.3 billion in 2025. Interest expense rose from $361 million to $1.229 billion.

Backlog says the customer promised to show up.

Interest expense says the bank already did.

The AI infrastructure owner has to live between those two sentences.

The Cloud Owner Owns the Asset

Amazon, Microsoft, Alphabet, and Meta are not Exodus. They have the scale to carry a mistake longer.

Scale buys duration.

The hyperscalers are buying the assets before the token economics have stabilized. They carry the capex, power, leases, financing, grid interconnection delays, construction risk, chip-cycle risk, and depreciation schedule. Customers get cheaper AI after the price cut. Labs get an adoption story. The cloud owner keeps the machine.

The income statement can stay calm because depreciation is spread over years.

Free cash flow has less patience.

Straight-line depreciation is a strange fit for AI infrastructure. Straight-line depreciation assumes the asset loses value smoothly. A GPU can have a short premium life and a longer residual life. The first part can fall quickly. The second part can linger for years.

Amazon's 2024 annual report gives the accounting lever. Amazon first increased the estimated useful life of servers from five years to six years effective January 1, 2024. That lowered depreciation and amortization expense by $3.2 billion and increased net income by $2.5 billion for 2024. Then Amazon said it would reduce the useful life of a subset of servers and networking equipment from six years to five years effective January 1, 2025, with an expected operating-income reduction of about $0.7 billion in 2025. It also recorded $920 million of accelerated depreciation and related charges for early retirements.

The chip does not care what useful life the accountant assigned to it.

The Frontier Lab Owns the Commitment

Frontier labs can inherit the GPU problem without owning the GPU.

The inheritance runs through compute commitments.

The cloud provider owns the depreciating asset. The frontier lab owns the take-or-pay problem.

During the boom, that commitment looks like access to scarce compute. After the pop, it looks like a fixed-cost trap. The lab still has to sell enough API usage, enterprise seats, agents, and consumer subscriptions to pay for capacity priced in a different market.

The cloud provider points to contracted backlog.

The lab points to falling token costs.

The contract sits between them.

If market-clearing inference prices fall faster than the contract price of compute, the risk was never removed. It was delayed.

Then the choices get ugly.

The lab can renegotiate. It tells the cloud provider the old capacity price no longer matches market economics. The provider can insist on the contract, but forcing the lab into insolvency may be worse than restructuring the deal.

The lab can break or default. Payment gets delayed. Capacity gets abandoned. The dispute becomes a legal and financial workout. The pain moves back to the cloud provider, who still owns the servers, the data-center obligations, the power contracts, and the debt.

The lab can try to fill the commitment with cheaper usage. Discount enterprise seats. Push API consumption. Bundle more AI into subscriptions. Route aggressively to smaller models. Sell longer agentic workflows. That protects utilization for a while. It can also make the pricing problem worse.

The first stage of the AI boom was about who could get enough GPUs.

The aftermath will be about who is stuck paying for them.

What Breaks After the Pop

The aftermath is a world where AI becomes normal infrastructure. Normal infrastructure has budgets, procurement rules, usage caps, gross-margin targets, depreciation schedules, debt covenants, and angry CFOs.

Consumer AI gets rationed first. The $20-per-month plan starts looking absurd when users expect frontier reasoning, coding agents, deep research, long context, image generation, video, voice, and tool execution inside one subscription. The consumer surface stays because it is distribution and brand power. The expensive parts get carved up: fewer frontier calls, stricter caps, slower queues, higher tiers, credits instead of unlimited use, more routing to smaller models, and premium pricing for agents, video, long context, and high-reasoning tasks.

Developer tools get less casual. The API rate card was already a fiction. The application layer added another one through bundles, multipliers, plan limits, and throttles. After the subsidy unwind, frontier calls stop feeling like tap water. Products that call the best model for every task inherit the subsidy. When the subsidy moves, the product margin moves with it.

Enterprise becomes the escape hatch. Frontier labs will push harder into enterprise because enterprise can absorb explicit pricing. Procurement can sign annual commitments. Compliance, security, admin controls, connectors, support, and data boundaries become the packaging around expensive inference. Enterprise gives the lab a better place to allocate, negotiate, and hide the cost.

Infrastructure owners defend backlog. Cloud providers and GPU lessors will point to committed contracts, multi-year demand, customer concentration, and residual value. Some of that defense will be valid. The machines can still run smaller models, batch inference, embeddings, internal workloads, lower-priority traffic, and enterprise jobs. The residual value is real. The premium-cost recovery is the issue.

The survivors get stronger. Companies with distribution, routing discipline, proprietary demand, low cost of capital, and control over more of the stack can absorb the shock. Weaker labs get acquired, merged, restructured, or turned into feature suppliers. The technology continues. The cap table changes.

Aftermath

When I bought Exodus, I thought I was buying the infrastructure layer of the internet.

In a narrow sense, I was right. The world needed data centers. The internet kept growing. Someone eventually used the buildings. Someone eventually made money from that kind of infrastructure.

The mistake was assuming the original investor owned the future.

AI may end the same way. The tools get better. The models get cheaper. Enterprises automate more work. Customers get more for less. The infrastructure survives in some form because the world still needs compute.

The fight is over the cost basis.

Who paid peak price for the machines? Who signed the take-or-pay contract? Who promised investors that backlog was safety? Who told customers that intelligence could get cheaper forever while the data-center bill waited quietly in the back?

The bill comes due there.

The future still arrives.

The original investors may never own it.

Sources