AI Tax and Subsidy

People think AI is making us more productive. Every tech conference, every earnings call, every startup pitch deck repeats the same mantra: AI is a transformational productivity tool that will reshape work, unlock human potential, and drive economic growth.

They're wrong.

What we're experiencing isn't productivity. It's a massive VC-funded subsidy masquerading as technological progress. And when the subsidy ends—and it will end—we're going to discover that most AI deployments are economic value destroyers, not creators.

This isn't cynicism. It's math.

The Economics of AI: A $5,000 Problem Sold as a $100 Solution

Here's how AI economics actually work in 2026:

You pay $100/month for an AI application. Let's say it's a B2B SaaS tool that uses AI to automate customer support or generate marketing content.

That company burns approximately $500/month on compute costs from AWS or Azure to serve you. They're losing $400 per customer per month, but they're okay with it because they're "building market share" and "demonstrating PMF to investors."

But it gets worse.

Amazon and Microsoft are burning roughly $5,000 per month to acquire GPU capacity from NVIDIA to provide that compute power. They're eating a 10x loss on every dollar of compute they sell.

The only reason you pay $100 instead of $5,000 is that venture capital is absorbing a 50x cost difference. This isn't a sustainable business model. It's a transfer payment from Sand Hill Road to end users, mediated through a complex supply chain.

When I explain this to execs, they usually say: "But costs will come down as models get more efficient and GPU supply increases."

Maybe. But that's not what the incentive structure suggests.

Why Costs Won't Fall Fast Enough

The AI cost curve is fighting against three powerful forces:

1. The Capability Treadmill

Every time models get cheaper, companies demand more capability. GPT-3.5 got cheap, so everyone moved to GPT-4. GPT-4 will get cheap, and everyone will move to whatever comes next. Claude Opus costs more than Sonnet, but companies pay it because they want the marginal improvement.

This is the same pattern we saw with cloud computing. AWS gets cheaper, but companies just consume more compute. The cost savings get competed away through feature expansion.

2. The Context Window Arms Race

Remember when 4K tokens was a long context window? Now we're at 200K+. Every doubling of context window roughly doubles compute costs. As AI applications get more sophisticated—ingesting entire codebases, processing multiple documents, maintaining long conversation histories—context requirements explode.

More context = more compute = higher costs.

3. The Multi-Modal Explosion

Text was expensive. Then we added images. Now we're adding video, audio, and real-time processing. Each modality multiplies the compute requirements.

The startups building AI video generators aren't just burning money on compute. They're burning industrial quantities of money. Some are spending $10-20M per month on GPU costs alone, pre-revenue.

So no, costs aren't coming down fast enough to save the economics. If anything, the capability demands are growing faster than the efficiency gains.

The Individual Illusion: Why You Think You're Productive

At the individual level, AI feels incredibly productive. I use Claude to draft documents, analyze data, and think through complex problems. It genuinely saves me time.

But here's the mental framework error: I'm confusing personal time savings with economic productivity.

Real productivity means: Output value generated / Resources consumed

If the subsidy disappeared tomorrow and my Claude subscription became $300-400/month to reflect true costs, would I still subscribe? Maybe for some use cases. Definitely not for others.

That "maybe" is the entire problem.

The AI productivity narrative works when someone else is paying the real cost. It falls apart when you have to pay it yourself.

The Enterprise Reality: Where AI Goes to Die

Now scale this individual illusion to enterprises and governments, and the economics get truly horrifying.

I've run AI implementations for both startups and large organizations. The difference isn't just speed; it's fundamental economics.

What's the difference? Why did the exact same technology succeed in one context and fail in another?

The Internality Problem: Organizations Optimizing Against Themselves

Traditional economics focuses on externalities—costs imposed on others. Pollution is an externality. Secondhand smoke is an externality.

But there's another failure mode: internalities. These are costs people impose on their future selves through bad decisions or lack of self-control.

Smokers understand intellectually that cigarettes are bad. But they start anyway because teenagers don't think about their 50-year-old selves. Then they can't quit because addiction overrides rational planning.

Large organizations have massive AI internality problems.

Here's what happens:

Phase 1: Resistance

AI is identified as strategically important. Multiple stakeholders claim ownership. Each department wants control but none want accountability. Governance processes designed for software purchases are applied to AI. Projects get stuck in "alignment" for months.

Phase 2: Pilot Purgatory

After 6-12 months, a small pilot is approved. Budget is constrained to minimize risk. Pilot is designed to prove value before scaling. But the pilot is too small to generate meaningful results. And it's measured on the wrong metrics.

Phase 3: The Failure Trap

Pilot shows "promising" but not "transformative" results. Stakeholders disagree on whether to continue. More analysis is commissioned. Original champions get promoted or leave. New stakeholders want "their" approach. Project dies quietly.

This pattern repeats across 90% of enterprise AI initiatives. And it's not because the technology doesn't work. It's because large organizations have governance structures that are optimized for preventing bad decisions rather than making good decisions.

The internality is this: Organizations know AI is strategically important. But their decision-making processes make it nearly impossible to deploy AI effectively. They're optimizing against their own long-term interests.

The Hidden Cost Structure: What Enterprises Actually Pay

Let's break down what a large enterprise actually spends on AI adoption:

Now ask: What's the return?

For most enterprises, it's murky at best. They can point to efficiency gains ("our customer service team handles 20% more tickets"), but they can't demonstrate clear ROI.

Why? Because the productivity gains are real but small, while the costs are massive and growing.

And remember: those software/API costs are artificially low because of VC subsidies. If you remove the subsidy, the economics get worse by an order of magnitude.

The Mental Framework: Externalities vs. Internalities

To understand where AI policy needs to go, you need to understand both externalities and internalities.

Externalities are costs imposed on others: Biased hiring algorithms that discriminate against protected groups. Surveillance AI that erodes privacy for entire populations. Social media algorithms that optimize for engagement at the cost of societal cohesion. Job displacement that creates social costs (unemployment, retraining, safety net).

Internalities are costs organizations impose on their future selves: Adopting AI too slowly and falling behind competitors. Adopting AI too quickly without proper governance and failing catastrophically. Building dependency on subsidized services that become unaffordable. Reorganizing around AI capabilities that don't actually deliver value.

Good policy needs to address both.

Right now, we're failing on both dimensions. We're allowing massive externalities to accumulate (algorithmic bias, privacy erosion, labor displacement) while simultaneously enabling internalities (organizations making bad long-term decisions based on artificially cheap AI).

The VC Subsidy as a Massive Externality

Here's the key insight: The VC subsidy itself is a form of externality—a cost being dumped on the future.

When VCs fund money-losing AI companies, they're not being charitable. They're making a calculated bet that some of these companies will achieve monopoly/oligopoly positions. Once entrenched, they can raise prices to profitable levels. Customers will be locked in and unable to leave. VCs will extract their returns during this transition.

But this creates a systemic externality. Organizations are making strategic decisions (reorganizing, retraining, building dependencies) based on artificially cheap AI. When prices rise to sustainable levels, they'll face a brutal choice: pay the real cost and destroy their economics, rip out AI and lose the capabilities they've built around it, or shut down entirely.

The companies that moved slowly and cautiously will actually be better positioned than the early adopters. The laggards won't have built expensive dependencies on subsidized services.

This is the opposite of how technology adoption usually works. Usually, first-movers win. In AI, first-movers are building technical debt on borrowed money.

Why 90% of AI Pilots Fail

The failure rate isn't a bug; it's a feature of the underlying economics.

AI pilots fail for three primary reasons:

1. The Value Isn't There Yet

Despite the hype, AI is still relatively narrow. It's great at specific tasks (text generation, image recognition, pattern matching) but struggles with complex reasoning across domains, tasks requiring deep contextual understanding, situations where errors are costly, and problems requiring true creativity or judgment.

Most enterprise use cases fall into these categories. So pilots show marginal improvements, not transformational change.

2. The Integration Costs Are Underestimated

Every AI pilot requires data preparation and cleaning, integration with existing systems, workflow redesign, user training and change management, and ongoing monitoring and refinement.

These costs are typically 5-10x the software costs. And they're pure expense; they don't scale, they don't compound, they just consume resources.

3. The Organizational Antibodies Are Strong

Organizations are optimized for stability, not change. Every AI implementation threatens someone's job, budget, or political position. The resistance is rational from an individual perspective, even if it's destructive from an organizational perspective.

So pilots get slow-rolled, starved of resources, or measured on impossible metrics. The failure is baked in from the start.

The Policy Imperative: Tax and Subsidy as Steering Mechanisms

If we let the market self-correct, we're looking at a catastrophic adjustment when VC funding slows and AI costs rise to sustainable levels. Enterprises and governments will be blindsided.

We need policy intervention now. And we have two primary levers: tax and subsidy.

The goal isn't to pick winners or losers. It's to correct the market distortion and align AI incentives with long-term societal value.

The Subsidy Framework: Where AI Creates Genuine Public Value

Government should subsidize AI deployment where it generates positive externalities—where social benefit exceeds private return.

Healthcare AI: Cancer Detection

Imagine an AI system that improves early cancer detection rates by 20%. The private return to the hospital is marginal (they get paid for scans regardless). But the social return is massive: lives saved, reduced treatment costs (early detection is cheaper), increased productivity from healthier population, and reduced suffering.

This is a textbook case for subsidy. Government should pay for deployment in smaller hospitals and clinics that couldn't otherwise afford it.

Education AI: Personalized Tutoring

An AI tutor that adapts to individual learning styles could dramatically improve educational outcomes, especially for disadvantaged students. The private return is limited (parents can't pay much). But the social return is enormous: better-educated workforce, reduced inequality, higher future tax revenue, and lower social costs (crime, welfare).

Subsidize deployment in public schools and underserved communities.

Climate AI: Modeling and Optimization

AI for climate modeling, grid optimization, and resource management generates massive positive externalities. The private return is often low or non-existent. But the social value is existential.

Heavy subsidy, broad deployment.

Government Services: Bureaucratic Automation

This is where subsidy makes the most sense. Government has enormous amounts of routine work that AI can automate: processing applications, answering citizen inquiries, analyzing regulations for compliance, and managing public records.

The private sector won't solve this (no profit motive). But the social value is significant: lower taxes (fewer employees needed), faster service (citizens wait less), better allocation of human workers to complex cases, and modernized government that actually works.

This is how you modernize a government in 90 days instead of 10 years. You subsidize AI deployment for routine work and redeploy humans to judgment-intensive tasks.

The Tax Framework: Where AI Creates Social Harm

Tax should be deployed where AI creates negative externalities—where private gain comes at social cost.

Social Media AI: Engagement Optimization

Facebook, TikTok, and YouTube use AI to optimize for engagement. The private return is massive (more engagement = more ads = more revenue). But the social cost is also massive: mental health damage (especially in teens), political polarization, misinformation spread, and erosion of shared reality.

Tax this heavily. Make companies internalize the social cost they're externalizing.

The tax structure could be simple: revenue from AI-driven engagement minus documented social value. If you can't prove your algorithm creates societal benefit, you pay the tax.

Surveillance AI: Privacy Erosion

AI-powered surveillance creates a massive negative externality: the erosion of privacy and civil liberties. Whether it's governments tracking citizens or corporations tracking consumers, the social cost is real.

Tax AI surveillance systems based on number of people surveilled, sensitivity of data collected, duration of retention, and scope of inference.

Make privacy erosion expensive. Force organizations to internalize the cost they're imposing on society.

Automated HR: Bias and Discrimination

AI hiring systems have been repeatedly shown to perpetuate and amplify bias. The private return is positive (cheaper than human recruiters). But the social cost is discrimination at scale.

Tax AI HR systems unless they can demonstrate regular bias audits, transparency in decision-making, accountability for errors, and measurably better outcomes than human processes.

If you can't prove your system is better than humans, you pay for the risk you're creating.

Job Displacement Without Transition

When AI automates jobs, companies capture the productivity gains while workers bear the adjustment costs. This is a classic externality.

Tax AI deployments based on job displacement, with exemptions for retraining programs funded by the company, gradual transition (not mass layoffs), creation of new roles requiring human judgment, and demonstrated productivity sharing with workers.

The goal isn't to prevent automation. It's to make companies internalize the transition costs they're currently externalizing.

The Tiered System: Carrots and Sticks Based on Impact

The smartest policy approach is a tiered system that combines tax and subsidy:

The tier determines your tax/subsidy rate. The specifics would vary by sector and use case, but the framework is clear: align private incentives with social value.

The Governance Challenge: Who Runs This?

AI policy is too complex for any single department. You need Treasury for tax design and collection, subsidy disbursement, and fiscal impact analysis. Tech Regulator for technical standards, audit methodologies, and certification processes. Labor Department for workforce impact monitoring, transition assistance, and retraining programs. Economic Policy Council for coordination across departments, strategic direction, and international alignment. Academic/Independent Bodies for research on impact, unbiased analysis, and public transparency.

But critically, you need a central coordinating body—call it an AI Economic Council—that has the authority to set strategic direction, resolve conflicts between departments, adjust policy based on evidence, report directly to political leadership, and maintain independence from industry capture.

The council should be staffed by people who understand both AI technology and economic policy. Not pure technologists (who undervalue social impact). Not pure economists (who undervalue technical constraints). But people who can think across both domains.

International Coordination: Why This Can't Be One Country

Here's the problem: AI is globally competitive. If the US implements smart tax/subsidy policy but China and Europe don't, we risk companies moving AI operations to lower-tax jurisdictions, brain drain to countries with looser regulation, competitive disadvantage for US firms, and inability to capture the benefits of good policy.

This requires international coordination. Not full harmonization (different countries have different values), but at least common frameworks for measuring AI impact, minimum standards for harmful applications, information sharing on what works, and coordination on subsidies to avoid races to the bottom.

The model is climate policy: national implementation with international coordination.

Small countries can actually win here. If you're Estonia, Singapore, or Israel, you can move faster than the US or China. Implement smart policy, attract AI companies doing genuine public good, and become a hub for high-value AI work.

Size isn't destiny. Speed and smart policy are.

The Productivity Paradox: Why Cheap AI Might Be Worse Than Expensive AI

Here's a counterintuitive insight: The VC subsidy making AI cheap might actually be harmful to productivity.

When something is cheap, we use it wastefully. When something is expensive, we use it carefully.

If AI cost its true economic price ($5,000/month instead of $100), organizations would deploy it only where the value genuinely exceeds the cost, invest more in making those deployments successful, build sustainable business models instead of dependency traps, and focus on high-value use cases instead of nice-to-haves.

The subsidy creates moral hazard. It encourages wasteful deployment, unsustainable business models, and hollowed-out organizations dependent on services they can't afford.

Raising AI prices to sustainable levels might actually increase productivity by forcing better allocation of resources.

The Coming Reckoning: What Happens When Subsidies End

The VC subsidy can't last forever. At some point—maybe 2-3 years, maybe 5-7 years—the money will tighten and AI companies will need to charge sustainable prices.

When that happens:

Scenario 1: The Soft Landing

Prices rise gradually. Efficiency improvements offset some of the increase. Organizations adjust and maintain critical AI deployments. Less critical uses are eliminated. We end up with more focused, higher-value AI adoption.

Scenario 2: The Hard Crash

Prices spike suddenly. Organizations are locked into dependencies they can't afford. Mass elimination of AI tools. Productivity losses from removing systems people rely on. Economic disruption, job losses in AI sector. Regulatory backlash and knee-jerk policy responses.

Which scenario we get depends entirely on whether we implement smart policy now.

If we guide AI adoption toward genuine value creation, we get the soft landing. If we let the subsidy bubble inflate further, we get the crash.

The Bottom Line: What Actually Matters

AI's productivity boom is largely a VC-subsidized illusion. We're burning capital to create the appearance of progress.

Real productivity means generating more value than you consume. Right now, AI is doing the opposite at scale.

But this doesn't mean AI can't be productive. It means we need to build the economic and policy infrastructure to ensure it actually is.

That requires honest accounting of AI's true costs and benefits, smart subsidies for applications with positive externalities, appropriate taxes for applications with negative externalities, coordinated governance across departments and countries, and long-term thinking that prioritizes sustainable value over short-term hype.

The organizations winning with AI aren't the ones deploying it everywhere. They're the ones deploying it strategically, where the value genuinely exceeds the cost.

The countries that will win the AI race aren't the ones with the most AI companies or the biggest compute clusters. They're the ones that align AI incentives with long-term societal value.

This is applied AI economics for people who actually run things. Not hype. Not fear. Just the math.

And the math says: We're building a productivity bubble on borrowed money. The question isn't whether it will pop. It's whether we'll have the wisdom to deflate it gradually, or the foolishness to let it explode.