The Biggest Threat to AI Investment Is AI Itself: How $700 Billion in Infrastructure Could Be Wiped Out by the Technology It Was Built to Run

Rapid algorithmic improvement poses the central threat to current AI infrastructure investment, because the same models being trained inside today's data centres are actively discovering more efficient architectures that require far less compute. A system optimised for transformer-scale workloads can become stranded capital within years if successor architectures demand fundamentally different hardware configurations.

There is a version of this story that the technology industry tells itself, and it goes like this: AI infrastructure spending is large, the returns will take time, but the demand is real and the economics will eventually follow. It is the same story told during the cloud buildout, and that one worked out. It is not a stupid story. But it is missing something important.

The most underexamined risk in the current AI investment cycle is not that demand fails to materialise. It is that AI will innovate the infrastructure underpinning that demand out of existence before the depreciation schedules complete their first half. The technology being built on top of these data centres is the same technology most likely to render them architecturally obsolete ahead of schedule. AI is not a passive consumer of compute. It is an active participant in a process that is recursively attacking the cost structure that justifies the capital being deployed on its behalf. No previous technology wave has done this. Railways did not invent teleportation. Oil companies did not discover infinite free energy. Telecoms did not build communication without bandwidth. But AI models are actively reducing token cost, model size, inference requirements, training requirements, and the hardware footprint required to do useful work. The system is attacking its own pricing structure, and the people financing the system are not fully accounting for that.

1. The Capital at Stake

The numbers involved are large enough that they deserve to be stated precisely rather than gestured at.

According to first-quarter 2026 earnings compiled by the Financial Times, Google, Amazon, Microsoft, and Meta collectively plan to spend $725 billion on capital expenditure in 2026, up 77% from last year’s record $410 billion. The breakdown per company is as follows.

Capital intensity has now reached 45 to 57 percent of revenue at the four largest hyperscalers, levels that are historically unprecedented. Approximately 75 percent of that aggregate spend, roughly $450 billion, is directed specifically at AI infrastructure. To fund this buildout, hyperscalers raised $108 billion in debt during 2025 alone, with projections suggesting $1.5 trillion in total debt issuance over the coming years. This is not equity risk at the margin. It is leveraged risk, with the time pressure that leverage implies.

The accounting assumption embedded in all of this is orderly five-year depreciation for compute hardware and significantly longer for buildings and power systems. That assumption is the foundation on which every financial model in this cycle is built. It is also, increasingly, fiction.

2. The Inference Price Collapse That Nobody Wants to Model

The single most important data series for evaluating AI infrastructure economics is not GPU shipment volumes or data centre power consumption. It is the price of inference, because inference revenue is what the infrastructure is ultimately meant to generate.

That price has collapsed at a rate with very few historical precedents in any technology market.

Inference costs for a model matching GPT-3.5 performance dropped from $20 per million tokens in November 2022 to $0.07 in October 2024, a 280 times decrease in roughly two years. LLM inference costs have fallen faster than nearly any computing commodity in history, with per-token prices declining between 9 and 900 times per year depending on the performance benchmark, and Gartner forecasting a further 90 percent cost reduction by 2030. A workload that cost $10,000 per month in 2023 now runs for under $200.

By April 2026, the price collapse has continued: Google’s Gemini Flash-Lite has set a new floor at $0.25 per million input tokens, and between early 2025 and mid-2026 alone, per-token costs for frontier models fell a further 60 to 80 percent across every major provider. The infrastructure being built today assumes that inference revenue will be sufficient to service the debt, cover the depreciation, and produce an acceptable return on equity. The price of inference is declining at a rate that makes that assumption structurally fragile, and it is declining not because the hyperscalers are choosing to compete on price, but because the underlying technology is making expensive inference progressively harder to justify.

3. The Economic Chain Reaction

The causal mechanism here is worth making explicit, because the argument is sometimes dismissed as speculative when it is actually a description of a process already in motion.

AI infrastructure spending assumes scarcity economics. Scarcity economics assume that expensive inference persists, because the models are large, the hardware is specialised, and the expertise required to operate it is concentrated in a small number of providers. Model efficiency is improving exponentially, driven by better training techniques, architectural innovations like mixture of experts, and increasingly by AI-assisted model design itself. Open source models are commoditising intelligence, making frontier-class capability available outside the closed API ecosystem. As open source quality converges with proprietary quality, inference pricing collapses, because enterprises can self-host rather than pay cloud rates. When inference margins compress, the revenue assumptions underlying $700 billion in annual infrastructure spending begin to fail. When revenue assumptions fail, the infrastructure becomes stranded.

That is not a theoretical chain. Each link in it is observable in current market data.

Simultaneously, hardware refresh cycles have already collapsed. What was once a five to seven year replacement standard in enterprise infrastructure has already compressed to two to three years in AI-specific data centre deployments. The depreciation schedules on infrastructure being built today are already misaligned with the replacement cycles being observed in practice. GPU price-performance has been doubling approximately every 2.5 years — and that sounds positive for the infrastructure owners. It is not. It means every two-and-a-half years, the infrastructure they just built is worth roughly half as much per unit of compute as what replaced it. The asset is depreciating faster than the accounting reflects.

4. The Open Source Collapse Scenario

The piece of this puzzle that receives the least rigorous treatment in mainstream financial analysis is the open source dimension, and it is the most structurally dangerous one.

DeepSeek’s V3 model, released in December 2024, claimed training costs of approximately $5.6 million compared to hundreds of millions for Western equivalents. It was competitive on benchmarks with models costing orders of magnitude more to produce. The market reaction was swift: Nvidia’s stock dropped nearly 17 percent within hours, representing the largest single-day market capitalisation loss in US corporate history at that point. DeepSeek’s R1 reasoning model launched at $0.55 per million tokens, at a time when OpenAI’s o1-preview, which had been released four months earlier with comparable capability, was priced at $15 per million tokens — a 97 percent discount for equivalent reasoning performance.

That was the warning. DeepSeek V4, released in April 2026, is the execution. V4-Pro carries 1.6 trillion total parameters — the largest open-weights model currently available — but activates only 49 billion of them per inference pass via a Mixture-of-Experts architecture trained on over 32 trillion tokens, with a native 1-million-token context window. It scores 80.6 percent on SWE-bench Verified, within 0.2 percentage points of Claude Opus 4.6, and is priced at $3.48 per million output tokens versus $25 for Claude. Both V4 variants are released under the MIT License and are downloadable from Hugging Face today. By activating only a fraction of parameters per forward pass, V4-Pro reduces compute per token by 60 to 80 percent versus dense architectures, structurally undermining the premise that frontier inference at scale requires dense, high-memory accelerators in massive quantities. Enterprise token costs dropped 67 percent year-over-year through 2025 into 2026, according to the AI Cost Consortium’s latest report.

The structural implication is not DeepSeek specifically. It is what DeepSeek demonstrated: that frontier-class AI capability can be built with radically less capital than the Western hyperscale model assumes, released under an open source licence, and replicated by any organisation with modest technical capacity. By 2026, enterprises hosting their own open source model instances are reporting reductions in AI operational expenditure of 85 to 95 percent compared to closed API rates. The commoditisation of the foundational model layer is not a future risk. It is a present reality.

Meta’s Llama family, released under a permissive licence, has produced models that organisations can fine-tune on proprietary data and deploy on their own hardware. Mistral, Phi, Gemma, and a deepening ecosystem of distilled and quantised derivatives have collectively ensured that the question enterprises now ask is not whether they can afford proprietary API access, but whether they can justify continuing to pay for it. Every enterprise that answers that question by self-hosting is a customer the hyperscale inference business loses permanently.

5. ARM and the Architectural Challenge to the GPU Monoculture

The current hyperscale buildout is architecturally concentrated. It is built almost entirely around Nvidia GPUs and the CUDA software ecosystem. That concentration creates a single point of structural vulnerability: if a plausible architectural alternative emerges that is materially more efficient, the economic case for the existing infrastructure degrades rapidly.

That alternative now exists. In March 2026, ARM released its first ever in-house production chip after 35 years of operating exclusively as an IP licensing business. The Arm AGI CPU is not a reference design — it is fully productised data centre silicon built on the Neoverse V3 platform, packing 136 cores into a 300-watt thermal envelope at 2.2 watts per core, compared to 3.9 watts per core for Intel’s 128-core Xeon 6. It claims more than double the performance per rack compared to traditional x86 platforms, using a rack-first design approach that prioritises density and thermal management, with over 8,000 cores and 180 terabytes of low-latency memory per standard 36 kilowatt rack — and up to 45,696 cores per rack in liquid-cooled 200 kW configurations. The chip delivers 800 GB/s memory bandwidth via 12 DDR5 channels, 96 PCIe Gen6 lanes, and native CXL 3.0 support. Beyond the flagship 136-core part, 128-core and 64-core variants are planned.

Launch partners include Meta as lead co-development partner, alongside OpenAI, Cerebras, Cloudflare, SK Telecom, SAP, Rebellions, Positron, and F5: the same organisations that are simultaneously the largest customers of the infrastructure the ARM architecture may be positioned to displace. Commercial systems are orderable today from ASRock Rack, Lenovo, and Supermicro, with production beginning in H2 2026, volume ramp in 2027, and ARM projecting $1 billion in chip revenue by 2028 scaling to an estimated $15 billion as customer onboarding accelerates.

ARM’s CEO has argued that CPU core counts will ultimately become more important than chip counts at the data centre level. If that argument proves directionally correct, it reframes the value of hundreds of billions of dollars in GPU infrastructure currently being depreciated on five-year schedules. The broader data centre CPU landscape has diversified sharply in parallel: AWS Graviton4 is in production at 96 Neoverse V2 cores with a 192-core Graviton5 in development; AMD’s EPYC Venice brings 256 Zen 6 cores on TSMC 2nm with a claimed 70 percent generational performance improvement; and NVIDIA itself has unveiled the Vera CPU with 88 custom Olympus cores targeting agentic orchestration workloads. Every major silicon vendor is now building for inference and orchestration, not training. The GPU monoculture is not being displaced — it is being surrounded.

6. Apple Silicon and the Inference Problem

While ARM challenges the CPU architecture of the data centre, Apple Silicon is challenging something more fundamental: the assumption that serious AI inference requires a data centre at all.

Apple’s M5, built on third-generation 3 nanometre technology, delivers over four times the peak GPU compute performance for AI workloads compared to its M4 predecessor, and introduces an 18-core Neural Engine capable of handling over 50 trillion operations per second. The architectural innovation here is not incremental. Apple has embedded Neural Accelerators directly into each GPU core rather than treating the neural engine as a separate component, fundamentally changing how AI workloads distribute across the silicon.

The power economics are a direct challenge to the data centre model. A Mac Studio with an M4 Ultra draws approximately 60 watts under full machine learning load. An Nvidia RTX 5090 draws 450 watts for the GPU alone, with total system power exceeding 600 watts. A $3,500 M4 Max Mac can run Llama 3 70B better than a $1,600 RTX 4090 setup because Apple’s unified memory architecture makes the full memory pool available to both the neural engine and GPU compute cores simultaneously, with no PCIe bottleneck. A three-year total cost of ownership comparison puts a Mac Studio cluster at $16,000 against an A100 server at $43,000 and AWS at over $80,000 for equivalent inference workloads.

Every workload that migrates to on-device inference is a workload that will never enter the hyperscale revenue line.

7. AI Is Designing the Chips That Will Obsolete the Chips It Currently Runs On

This is the mechanism that receives almost no attention in mainstream coverage, and it is arguably the most structurally significant one.

The chips being purchased today at scale were designed largely by human engineers working within the constraints of existing tooling. The chips that will be available in two to three years will be designed with substantially more AI assistance, optimised for the specific workload profiles that have emerged since the current generation was specified, and manufactured on process nodes that have not yet reached production. Google has already used AI to design its Tensor Processing Units, with AI-assisted floorplanning reducing a process that previously took months to under six hours. This loop is now operating across the industry.

GPU price-performance has been doubling approximately every 2.5 years, and specialised AI chips have been accelerating this trend substantially. In the AI sector, hardware refresh cycles have already collapsed from the five to seven year standard of the early 2010s to two to three years, with software obsolescence now driving hardware obsolescence. The assets being depreciated over five years will face architecturally superior, radically more efficient alternatives within two. The accounting model reflects the world as it was, not the world as it is moving.

The hyperscalers are financing nuclear power stations while the technology is sprinting toward solar efficiency curves. The plants will still be standing when the economics of building them no longer make sense.

8. What If Demand Outruns Efficiency Gains?

This is the strongest rebuttal to the thesis, and it deserves a direct answer rather than being left unstated.

The bull case is that cheaper inference creates exponentially more usage, that agentic AI multiplies token demand by orders of magnitude as models chain workflows and call each other recursively, that video AI and robotics drive persistent inference demand at a scale that dwarfs current projections, and that AI becoming embedded in every software product means that even a dramatically cheaper per-token price generates more aggregate revenue than today’s expensive per-token price generates now. Jensen Huang’s framing of an “agentic AI inflection point” is essentially this argument.

It is not implausible. Total token consumption is rising faster than prices decline, because modern reasoning models loop and chain workflows in ways that burn far more tokens per request than earlier systems. The average enterprise AI budget has grown from $1.2 million per year in 2024 to $7 million in 2026. The paradox that the cost of intelligence is falling while the cost of deploying intelligence is rising is real, and it does create a partial offset.

The problem is timing and margin. Agentic demand at scale is a 2027 to 2030 story at the earliest. The infrastructure spending is a 2025 to 2026 story. The depreciation schedules start running now, against revenue assumptions that depend on a demand curve that has not yet materialised. And the efficiency gains are not waiting for demand to catch up: per-token inference prices are falling between 9 and 900 times per year depending on the benchmark, which means the margin available per unit of demand is compressing at a rate that even dramatic volume growth struggles to offset. The demand argument does not disprove the infrastructure risk. It modifies the probability distribution of outcomes, and leaves the tail risk largely intact.

9. The Stranded Asset Question

Credit analysts are already flagging infrastructure obsolescence as a material risk, with hyperscaler bond credit spreads widening during the first quarter of 2026 in a pattern consistent with markets beginning to price uncertainty about whether AI revenue can scale fast enough to justify the spending. Amazon’s free cash flow is projected to turn negative in 2026. Morgan Stanley expects hyperscaler debt issuance to exceed $400 billion. The buildout is increasingly debt-financed, meaning the capital base sitting underneath it is not purely equity risk.

The most honest historical comparison is the 1990s telecom fibre boom, which destroyed over $2 trillion in equity value, not because the demand for bandwidth was wrong but because the infrastructure was built ahead of the demand curve, at a cost structure that the eventual market price of bandwidth could not service. The AI infrastructure cycle shares that structural characteristic. The fibre was real. The demand was real. The economics were catastrophic for investors who got the timing wrong.

The difference, and it is a crucial one, is that in the fibre boom, the technology being deployed over the fibre did not actively work to make the fibre cheaper to replace. In the current AI cycle, the technology being deployed on the infrastructure is iteratively improving the efficiency of the models, the chips, and the architectures that run on it, in a recursive loop that is compressing the economic lifespan of the hardware faster than any previous infrastructure buildout has experienced.

10. So What Does This Mean?

The winners in this cycle are likely to be the organisations that own the things AI cannot commoditise: electricity generation and distribution, physical land with power and cooling, the fabrication capacity to produce the next generation of chips, and the distribution relationships that sit between the model and the end user. The hyperscalers themselves may survive and even thrive, but their returns on the current infrastructure cycle will depend heavily on whether agentic demand arrives before efficiency gains collapse their inference margins entirely.

The losers, in the scenario where the thesis is correct, are the organisations that have committed the most capital to the current architectural paradigm on the longest depreciation schedules, with the most leverage, at the highest cost per unit of compute. They are betting that the infrastructure they are building today will still be load-bearing in five years. Given that AI is designing the chips, compressing the models, building the open source alternatives, and collapsing the inference price curve, that is not a bet anyone should take at face value.

To make the stakes concrete: DeepSeek V4-Pro is open-weighted, MIT-licensed, running at 49 billion active parameters across 1.6 trillion total, delivering frontier-class performance at $3.48 per million output tokens. The ARM AGI CPU — the first production chip in ARM’s 35-year history — is in volume production with Meta and OpenAI as launch partners. Apple’s M5 Neural Engine handles 50 trillion operations per second in a consumer laptop. Inference costs have fallen 99.7 percent in three years and are still falling. Each of these is not a warning of what is coming. It is a description of what is already here.

The danger here is not that AI fails. It is that AI succeeds so quickly, and in such architecturally unexpected directions, that the infrastructure bet placed today is correct about the destination and catastrophically wrong about the route.

Sources:

• DeepSeek V4 Preview Release — DeepSeek API Docs
• DeepSeek V4: Features, Benchmarks, and Comparisons — DataCamp
• DeepSeek V4 Pro Benchmarks 2026 — BenchLM.ai
• Announcing Arm AGI CPU — Arm Newsroom
• Arm’s 136-Core AGI Chip Outpaces x86 in Data Centres
• Arm’s $15 Billion CPU Opportunity — Futurum Group
• CPUs are Back: The Datacenter CPU Landscape in 2026 — SemiAnalysis
• NVIDIA Blackwell: 5× H100 Performance at $40K per GPU
• Hyperscalers Hit $700 Billion in 2026 AI Spending Plans — Yahoo Finance
• LLM Pricing Collapse of 2026 — AI Magicx
• AICC Report: Enterprise Token Costs Drop 67% Year-Over-Year
• LLM API Pricing Comparison 2026 — BenchLM.ai