The Audacious Plan to Tax Every AI Computation and Build America’s First Sovereign Wealth Fund

A California billionaire has proposed what may be the most unusual tax idea in Silicon Valley history: charge a fee on every token processed by artificial intelligence systems and use the revenue to create a sovereign wealth fund for the American people. The proposal comes from Vinod Khosla, the venture capitalist and Sun Microsystems co-founder whose net worth Forbes pegs north of $10 billion, and it’s already generating fierce debate across the technology industry and in policy circles in Washington.

The idea is disarmingly simple. Every time an AI model processes a token — the basic unit of text, code, or data that large language models like GPT-4 or Claude consume and generate — a small tax would be levied. Fractions of a cent per token. Individually trivial. But collectively, given the hundreds of billions of tokens processed daily across the AI industry, the sums would be staggering. Khosla envisions funneling that revenue into a national sovereign wealth fund modeled loosely on Norway’s Government Pension Fund, which has accumulated over $1.7 trillion by taxing the country’s oil revenues and investing them on behalf of its citizens.

As Gizmodo reported, Khosla laid out his thinking in a series of public statements arguing that AI will generate enormous wealth — but that wealth risks concentrating in the hands of a few technology companies and their investors unless policymakers intervene early. His argument isn’t anti-capitalist. Far from it. Khosla is one of the most prolific venture investors in AI, with stakes in companies across the sector. He’s arguing, essentially, that the technology he’s helping to build will be so productive, so enormously profitable, that society needs a mechanism to distribute some of those gains broadly.

“AI is going to be the largest wealth creation event in human history,” Khosla has said. He wants ordinary Americans to have a stake in that outcome.

The mechanics of a per-token tax raise immediate questions. Who pays it? The companies running the models? The developers calling the APIs? The end users? Khosla has suggested the tax would fall on the AI providers — companies like OpenAI, Google, Anthropic, Meta, and others operating large-scale inference infrastructure. In practice, this would function something like a severance tax on natural resource extraction, except the resource being “extracted” is computational intelligence rather than oil or timber.

And that analogy is precisely the one Khosla wants people to hold in their minds. Norway didn’t let its oil wealth flow exclusively to petroleum companies and their shareholders. It captured a portion for the public. Alaska does something similar with its Permanent Fund, which pays annual dividends to every state resident from oil revenues. Khosla’s pitch is that AI tokens are the new barrels of oil — a resource whose value is immense, whose production is scaling exponentially, and whose benefits should be shared.

Not everyone is buying it.

Critics within the technology industry have responded with a mix of skepticism and alarm. Some argue that taxing tokens would amount to taxing thought itself — a computational process tax that would slow innovation, raise costs for consumers, and drive AI development offshore to jurisdictions without such levies. Others point out that the token is a poor unit of taxation because different models use tokens differently, and the relationship between tokens processed and economic value generated is wildly inconsistent. A token used to generate a haiku produces far less economic value than a token used to analyze a medical scan or optimize a supply chain.

There’s also a definitional problem. What counts as a token? The term means different things across different model architectures. Tokenization schemes vary. Some models are more token-efficient than others. A tax per token could inadvertently penalize less efficient models while giving an advantage to companies that have optimized their tokenization — creating perverse incentives that have nothing to do with the policy’s stated goals.

But Khosla’s proposal doesn’t exist in a vacuum. It arrives at a moment when policymakers at every level of government are grappling with how to respond to AI’s rapid commercial deployment. The Trump administration has largely favored a deregulatory approach, rolling back Biden-era executive orders on AI safety and signaling that the United States should prioritize AI dominance over AI regulation. Congressional action has been slow and fragmented, with dozens of AI-related bills introduced but few advancing through committee.

The sovereign wealth fund concept, however, has been gaining traction independent of AI. Several U.S. lawmakers have floated proposals in recent years for a national sovereign wealth fund, though none have gained significant legislative momentum. President Trump himself mentioned the idea in passing during his 2024 campaign, though without specifics on funding mechanisms. Khosla’s contribution to the debate is to connect two separate policy conversations — what to do about AI’s economic disruption and how to build long-term national wealth — into a single proposal.

The scale of AI inference is worth pausing on. OpenAI alone reportedly processes hundreds of billions of tokens per day across its ChatGPT and API products. Google’s Gemini models, Anthropic’s Claude, Meta’s Llama-based deployments, and dozens of smaller providers add enormously to that total. Industry analysts estimate that global AI inference compute is growing at roughly 10x per year. Even a tax of a tiny fraction of a cent per token — say, one-hundredth of a cent — could generate billions of dollars annually at current volumes. And volumes are only going up.

Khosla has been characteristically blunt about the political dynamics. He’s acknowledged that the technology industry will fight the proposal. He’s also argued that the industry should welcome it as an alternative to far more aggressive interventions — breakups, heavy-handed regulation, or windfall profit taxes — that could come later if the public perceives that AI’s benefits are flowing only to billionaires and big tech shareholders. In his framing, a token tax is the moderate option. The market-friendly option. The option that lets companies keep innovating while ensuring the public gets a cut.

This framing has drawn comparisons to the debates over taxing financial transactions. The so-called Tobin tax — a small levy on currency trades proposed by economist James Tobin in the 1970s — has been debated for decades. Proponents argue it would generate revenue and reduce speculative trading. Opponents say it would reduce market liquidity and push trading to unregulated venues. The European Union attempted a financial transaction tax starting in 2013; a decade later, only a handful of member states have implemented anything resembling it. The history suggests that per-unit taxes on high-volume digital transactions are easier to propose than to implement.

So where does this leave the proposal? In the near term, it’s more conversation starter than legislation. No bill has been introduced in Congress. No regulatory agency has signaled interest. Khosla himself seems to understand that the idea’s value right now is primarily rhetorical — a way to shift the Overton window on AI taxation and force the industry to engage with questions about wealth distribution before the political environment turns hostile.

There are also deeper philosophical questions embedded in the proposal. If AI systems eventually perform a significant share of the cognitive labor currently done by humans — and many in the industry believe this is a matter of when, not if — then the economic output of those systems represents something genuinely new. It’s not quite capital income. It’s not quite labor income. It’s something else: machine cognition generating economic value at scale, with minimal human involvement per unit of output. Existing tax frameworks weren’t designed for this. A token tax is one attempt to create a new framework. Whether it’s the right one is an open question.

Khosla’s track record gives the proposal more weight than it might otherwise carry. He co-founded Sun Microsystems in 1982, was a general partner at Kleiner Perkins, and launched Khosla Ventures in 2004. His firm has invested in dozens of AI companies. He’s not an outsider lobbing grenades at an industry he doesn’t understand. He’s an insider arguing that the industry he helped build needs to think about its social contract before someone else defines it for them.

The sovereign wealth fund piece of the proposal also deserves scrutiny. Norway’s fund works because it’s governed by strict rules about investment and withdrawal, insulated from short-term political pressures. Alaska’s Permanent Fund has been more politically contentious, with periodic fights over dividend levels and fund management. Creating a U.S. sovereign wealth fund would require answering hard questions about governance, investment mandates, distribution mechanisms, and political independence. Would Congress be able to resist raiding the fund during budget fights? History suggests not.

And yet the underlying intuition — that AI will create enormous wealth and that some mechanism should exist to share it — resonates with a broad swath of the American public. Polls consistently show that Americans are anxious about AI’s impact on jobs and inequality. A proposal that says “you’ll get a check from the AI economy” has obvious political appeal, even if the implementation details are fiendishly complex.

The coming months will likely see more proposals along these lines. As AI companies move toward IPOs, as inference costs continue to fall while volumes explode, and as the economic impacts of AI become more visible in labor markets, the pressure on policymakers to respond will intensify. Khosla’s token tax may not be the answer. But it’s asking the right question: who benefits when machines start thinking for a living?

For now, the proposal sits in that uncomfortable space between visionary and impractical. A billionaire venture capitalist proposing to tax the industry that made him rich. A market-based solution to a problem that markets alone may not solve. An idea that sounds radical but might, in retrospect, look like the moderate path not taken. The technology industry has a long history of dismissing policy proposals until they become law. This one is worth watching.