>_TheQuery
← All Articles

The VC Subsidy Behind Cheap AI Will Not Last

By Addy · March 15, 2026

Perplexity CEO Aravind Srinivas recently amplified a post arguing that AI is automating routine coding tasks and that computer science is heading back to its roots: math, physics, and logical thinking. The post crossed a million views on X. The argument is credible. The direction is probably right.

But there is a question underneath it that nobody is asking loudly enough.

Who is paying for all of this?

The Subsidy Nobody Talks About

Every time you send a complex query to Claude, GPT-5, or Gemini, the AI lab is losing money on the transaction.

That is not speculation. OpenAI is projected to burn $14 billion in 2026, up from $8-9 billion in 2025. The company is projecting cumulative losses of $115 billion through 2029. Anthropic's margins swung from -94% in 2024 to roughly +40% in 2025, but remain under pressure from inference costs that keep running higher than expected.

In February 2026, 90% of VC funding dollars went to AI startups. OpenAI and Anthropic alone captured 74% of all VC dollars globally. More than half of all venture capital is now flowing into AI companies.

The reason AI feels cheap, or free, is not that it is cheap. It is that someone else is paying the difference. Venture capital is underwriting the gap between what AI costs to run and what companies charge for it. The same way VC money subsidized your $4 Uber rides in 2015 and your $8 DoorDash deliveries in 2020.

We know how those stories ended.

The Millennial Lifestyle Subsidy, AI Edition

Uber raised billions. Rides were cheap. The subsidy normalized cheap rides as the baseline expectation. Then Uber needed to become profitable. Prices went up 40-60%. The users who had built their commute around $4 rides discovered that $4 was never the real price.

AI is running the same playbook, at a much larger scale.

The developers who are replacing routine engineering tasks with AI today are doing it at subsidized prices. The companies integrating AI into their workflows are building cost models around subsidized inference. The startups being told they can replace a software engineer with an AI agent are calculating that replacement at subsidized API rates.

Nobody knows exactly when the subsidy ends. The variables are real: chips are getting more efficient, models are getting smaller (as we covered in the Nemotron 3 Super and Qwen articles), and competition between labs is keeping prices artificially low. The direction of AI costs over the long run is probably down.

But "probably down eventually" and "definitely free forever" are not the same thing. At some point between now and the OpenAI IPO (Reuters reports it could value the company at $1 trillion), public market investors will demand margins. Losses of $14 billion a year do not survive an earnings call.

What Happens When Companies Actually Have to Pay

This is the question the "AI replaces engineers" conversation is missing entirely.

The argument goes: AI handles routine coding, so companies need fewer engineers for routine tasks. That is probably true today, at today's prices. The calculation looks different at 3x the current API cost.

A company that replaced three junior developers with AI agents at $500/month in API costs is making a clear financial decision. The same company facing $1,500/month, or $5,000/month for the same workload as usage grows, is running a different calculation. Some will still find it worth it. Some will not. The engineers they let go are not coming back.

The more important shift is at the enterprise level. TechCrunch surveyed 24 enterprise-focused VCs and the consensus was clear: enterprises are moving from experimenting with many AI tools to concentrating spend on fewer proven vendors that deliver measurable ROI. The era of "AI everywhere because everyone else is doing it" is ending. What comes next is "AI only where it demonstrably pays for itself."

That consolidation is healthy for the industry long term. It is not painless for the companies and the engineers who made decisions based on the subsidized version of AI economics.

The Variable Nobody Can Solve

How much will costs go up? Genuinely unknown.

The optimistic case: architectural improvements like sparse Mixture of Experts keep driving inference costs down faster than demand drives them up. Labs reach sustainable margins without significant price increases. The subsidy quietly ends and nobody notices because efficiency absorbed the gap.

The pessimistic case: inference costs are stickier than the efficiency gains because demand is growing faster than chips can handle it. Labs reach IPO, investors demand profitability, prices go up 2-3x. The companies that over-indexed on AI during the subsidy period face a reckoning.

The realistic case is somewhere between those two. AI costs will not go to zero. The current pricing is not permanent. The gap between "AI is cheaper than an engineer" and "AI is more expensive than an engineer" is smaller than most people building on current prices assume.

What This Means for the Engineers

The Perplexity CEO's amplified post is right about one thing: the engineers who survive the transition will not be the ones who memorized syntax. They will be the ones who understand systems, can reason about architecture, and can evaluate whether an AI agent's output is correct.

That skill set is genuinely more valuable than the one it replaces. The argument for learning math and physics over memorizing frameworks is real.

But the transition assumes the AI tools enabling that shift remain economically accessible. A $20/month developer subscription that replaces $50,000 in API costs is a different calculation than a $500/month subscription that replaces $150,000. Both might still be worth it. The math is different.

The engineers who are watching AI automate their routine tasks should probably be paying less attention to whether AI can do their job and more attention to whether the companies using AI to do it can afford to keep doing it at current prices.

That question does not have a clean answer. It is worth asking anyway.

Sources:

Previously on TheQuery: The Model That Thinks With 12B Parameters but Knows Everything a 120B Model Knows: the architectural shift that may be the optimistic case for why costs stay manageable.