Part 6 of 6 in the High Country Advocate’s investigation into AI infrastructure spending
We started by asking where $405 billion goes. We documented three existential problems getting worse, not better. We traced the money to infrastructure providers profiting massively while application companies lose billions. We showed ratepayers funding the buildout through higher electricity bills. We examined whether distributed computing could offer an alternative.
The evidence points to clear conclusions based on documented facts about money flows, technical constraints, and historical patterns.
The spending is real and accelerating. Big Tech allocated $405 billion for AI infrastructure in 2025, representing 94% of their operating cash flows. McKinsey projects $5.2 trillion needed by 2030. Goldman Sachs estimates $720 billion for grid upgrades alone. These aren’t projections—they’re funded commitments.
Infrastructure providers are profiting enormously. Nvidia generated $130.5 billion in revenue for fiscal 2025 at 70-78% gross margins—software-like returns on hardware products. The semiconductor supply chain profits from volume. Cloud providers markup AI compute costs. All get paid regardless of whether AI applications succeed.
AI application companies are bleeding cash. OpenAI lost $5 billion in 2024 on $3.7-4 billion in revenue—spending $2.25 to earn $1. The company projects $44 billion in cumulative losses through 2028. Anthropic lost over $5 billion on under $1 billion in revenue. Both expect profitability years away despite massive user bases.
The revenue gap is enormous. Total AI application revenue reaches $20-30 billion in 2025. Justifying $5.2 trillion in infrastructure by 2030 requires $2 trillion in revenue—100x growth in five years. This must happen while 95% of AI projects fail to deliver measurable impact, according to independent research from MIT, Gartner, Harvard Business Review, and RAND.
Ratepayers are funding the infrastructure. Electricity prices jumped 6-20% in regions with data center concentrations. PJM capacity market costs increased from $2.2 billion in 2022 to $16.1 billion in 2025, with data centers accounting for 63%. Bloomberg found wholesale prices up to 267% higher than 2020 near data centers. These costs transfer directly to consumer bills.
Distributed computing faces fundamental constraints. Light travels one foot per nanosecond—New York to California requires minimum 40 milliseconds round trip. AI model training requires microsecond coordination. The 1,000x latency difference breaks training loops. Edge computing works for specific applications but can’t replace centralized infrastructure for frontier AI.
Historical patterns are clear. Infrastructure providers profit regardless of application success. Most companies fail. It takes longer than expected. Value consolidates to survivors. The railroad boom and telecom bubble followed this pattern. The key difference: AI has ongoing operating costs—electricity, cooling, compute—that railroads and fiber didn’t require at comparable scale.
Four scenarios emerge from this evidence.
The soft landing that industry hopes for requires AI application revenue to grow 100x by 2030, reaching $2 trillion. The 5% of projects that succeed would need to drive spectacular productivity gains. Companies would achieve profitability on schedule. This has no historical precedent in technology adoption—enterprise software took decades to reach similar scales—and requires the 95% failure rate to reverse dramatically. Probability: Very low.
The most likely outcome matches historical patterns. Infrastructure providers continue profiting while AI application companies consolidate. Two or three major players survive. Most AI startups fail or get acquired. Revenue growth happens but takes 10-12 years, not 5. Nvidia, cloud providers, and the semiconductor supply chain capture most value. Ratepayers continue funding infrastructure through electricity bills. This is what happened with railroads and telecom—sell shovels in a gold rush, don’t mine gold. Probability: Medium-high.
A bubble pop would vindicate Michael Burry, who put $1.1 billion behind his thesis that $176 billion in AI earnings are overstated through depreciation accounting. Revenue growth doesn’t materialize. Investors lose patience after sustained losses. Massive write-downs follow. Data center capacity sits underutilized. Electricity rates stay elevated because infrastructure costs are sunk. The ingredients are present: excessive spending relative to revenue, optimistic projections, circular financing. Probability: Medium.
A distributed computing revolution would require overcoming fundamental physics constraints. No peer-reviewed research demonstrates viability for frontier AI. Jay Valentine’s claims remain unverified by independent technical review. Probability: Very low.
The evidence points most strongly toward consolidation with elements of correction. Infrastructure providers have already captured significant value and will retain most of it regardless of application outcomes. AI applications will consolidate to a few well-funded survivors. Most startups will fail. Revenue growth will happen but slower than required to justify current spending.
Some write-downs seem inevitable. Companies are depreciating AI chips over 5-6 years when useful lives may be 2-3 years. That accounting produces overstated earnings now and painful corrections later. But total collapse is less likely than painful adjustment. The infrastructure has real utility even if current economics don’t work. Survivors will eventually achieve profitability, just later and at lower levels than projected.
Ratepayers will continue bearing infrastructure costs. The investments are sunk. Utilities have regulatory approval for recovery. Future expansions will face more scrutiny as political pressure builds—Virginia and New Jersey’s governors won election partly on this issue—but existing costs are locked in.
For policymakers: Require data centers to fund infrastructure costs through direct charges rather than socializing expenses across all ratepayers. Job creation and tax revenue don’t justify electricity subsidies when consumers pay higher bills.
For investors: Infrastructure providers represent the safest positions—already profitable, pricing power, paid regardless of application success. AI application companies carry high risk. Most will fail. Watch depreciation accounting closely.
For businesses: The 95% failure rate applies broadly. Start small, measure carefully, focus on specific use cases with clear ROI. Avoid general “AI strategy” initiatives without defined outcomes. Plan for 3-5 year payback periods, not 6-12 months.
For consumers: Electricity bills will likely remain elevated. Pressure elected officials to stop subsidizing future data center expansion through ratepayer bills.
Nvidia has already made its money—$130.5 billion in fiscal 2025 at 78% margins. Cloud providers are collecting their premiums. The semiconductor supply chain is profiting from volume. These companies are paid regardless of what happens next.
AI application companies face $44 billion in projected losses through 2028, unit economics that don’t scale like traditional software, and revenue requirements that exceed historical adoption patterns. Some will survive. Most won’t.
The math requires either unprecedented revenue growth justifying $5 trillion in infrastructure, or the industry is building on speculation that won’t materialize. Current evidence—95% failure rates, 100x required growth, $2.25 spent per dollar earned—suggests the latter.
Someone will lose a lot of money. History suggests it won’t be Nvidia.
Sources: This article synthesizes evidence from Articles 1-5. Financial data from Nvidia earnings reports, OpenAI/Anthropic investor documents, Big Tech capital allocation filings, and Michael Burry SEC 13F filings. Technical research from MIT, Gartner, Harvard Business Review, RAND, and academic distributed computing studies. Electricity data from PJM Interconnection, Bloomberg, and state utility filings. For complete citations, see documentation in Articles 1-5.
Support Independent Local Journalism — High Country Advocate was created as a real alternative to regional media that too often silences dissenting voices while taking sides in the political struggle. Producing in-depth, unflinching reporting like this series and others is expensive: servers, editing, research time, and legal review all add up quickly. If these articles have informed you or given you new perspective, please consider supporting HCA with a paid subscription — every subscriber helps keep this reporting strong and independent. 