Jensen Huang Is Taking a Cut Now

When Compute Becomes a Financial Instrument, Taiwan’s Three Moves

I. July 2: The Shovel-Seller Takes Equity in the Gold Miners

On July 2, 2026, NVIDIA announced a new commercial model, rolled out globally: revenue sharing.

The rules are simple. AI startups no longer need to raise hundreds of millions to build data centers and buy GPUs. NVIDIA extends token credits——compute now, repaid later as a share of future revenue. NVIDIA CFO Colette Kress, in the official announcement, called the resulting income:

“a recurring, usage-linked earnings stream.”

Two launch partnerships were named:

• Sharon AI: deploying up to 40,000 Grace Blackwell GB300 GPUs. CEO James Manning called it “a pivotal moment in Sharon AI’s mission to deliver sovereign, large-scale AI compute infrastructure.”
• Firmus Technologies: building the DSX AI factory campus on Batam Island, Indonesia——planned to scale to 360 megawatts and house up to 170,000 GPUs.

International coverage named Baseten, Fireworks AI, and Together AI as likely next-wave participants. And one line in CNBC’s report deserves a second read: GPUs are now being likened to oil——and reportedly tied to futures contracts.

Translated into plain language:

In phase one of the gold rush, NVIDIA sold shovels——cash on delivery. Phase two has begun: NVIDIA no longer just sells shovels. It takes equity in the miners and skims every gram of gold dust.

One-time sales become rent. A hardware company becomes a platform. And a platform’s next move——every chapter of commercial history agrees——is always the same:

It becomes a bank.

II. Compute Financialized: Three Layers of Power, and Taiwan Stands on Only One

First, why this news matters more than it looks.

Over the past year I proposed the Compute Purchase Agreement (CPA)——bulk-purchase compute on long-dated contracts, modeled on the energy sector’s PPAs——in Musk Is Building 10 GW in Texas. What About Taiwan?, and the Strategic Compute Reserve (SCR)——forward purchases of 5 to 10 years of AI tokens, modeled on the Strategic Petroleum Reserve——in No Compute Sovereignty, No National Sovereignty.

Both essays shared one premise: compute is turning from equipment into a commodity, and from a commodity into a financial instrument.

On July 2, NVIDIA turned that premise into a product. Token credits are a line of credit. Revenue share is the interest. Future revenue is the collateral. This is a compute bank’s charter——it just isn’t framed on the wall yet.

Now split compute power into three layers:

The June 13 model directive was the allocation layer drawing its sword for the first time——Washington deciding, by national security order, who may use frontier models. The July 2 revenue-share model is the allocation layer’s second evolution——NVIDIA deciding, by credit review, who may use compute at all.

The first is political allocation. The second is financial allocation. Stack them together and you get the completed form of the sentence I wrote three weeks ago:

No compute sovereignty, no national sovereignty.

To which we must now append a second half:

Compute sovereignty now has a price——and an interest rate.

Pope Leo XIV’s encyclical warned that “the control of AI must not fall into the hands of a few.” Three weeks later, June 13 proved the few can decide who uses the models. Three weeks after that, July 2 proved the same few can decide who gets compute on credit, and at what cut.

When the compute bank opens for business, a nation that is merely a depositor or a borrower has mortgaged its AI industry to someone else’s credit committee.

Taiwan, as of today, has not even opened an account.

III. The Warning from Batam: Someone Else Is Executing the Archipelago Strategy

Of the two launch partnerships, the one that should keep Taipei awake at night is not Sharon AI’s forty thousand GB300s. It is Firmus choosing Batam Island, Indonesia.

Five weeks ago, when I proposed offshore satellite compute stations——local land and local power, matched with external capital and technology, sharing the resulting compute——Southeast Asia was on my candidate list. Some readers said it was far-fetched.

Now look at Batam: Indonesia contributes land and 360 megawatts, Firmus contributes operations, NVIDIA contributes GPUs and the financial structure. That is a satellite compute station. It just isn’t flying Taiwan’s flag.

Look also at Sharon AI’s language: “sovereign, large-scale AI compute.” Note carefully——this is not the vocabulary of a government white paper. It is the vocabulary of a sales deck. Sovereign compute is now a sellable product category. Sold to whom? To every country that has discovered it holds no seat on the algorithmic tier.

The contrast is brutal:

• Taiwan: ships the world’s GPUs, power supplies, cooling systems, and server racks——then debates whether to deploy compute overseas.
• Indonesia: fabricates not a single chip——and lands 170,000 GPUs in exchange for land and power.

We sit at the very top of the supply chain and the very bottom of the value allocation.

Batam is 20 kilometers from Singapore and 3,000 kilometers from Kaohsiung. But on the compute map, it now sits closer to the center of the AI era than any science park in Taiwan.

IV. Taiwan’s Three Moves

Every proposal from the previous two essays——the sovereign fund, the compute reserve, the Compute Embassy, bulk-purchase compute——still stands. But after July 2, the board has a new player and a new rulebook. The three moves below plug the existing proposals into NVIDIA’s new model:

Move 1: Be Firmus, Not Just a Customer

NVIDIA’s revenue-share model needs two kinds of partners: startups that consume compute (the borrowers), and operators that build AI factories (the branches).

The Firmuses of the world are claiming the second seat——assembling land, power, and facilities, racking NVIDIA’s GPUs, and taking a share of the recurring revenue.

That is the seat Taiwan must contest. The logic is direct: no country on earth is better equipped to operate AI factories——

• Taiwan’s flagship foundry and flagship power-electronics maker hold the two most expensive pieces of any AI factory;
• A certain systems-assembly group already operates major sites in Wisconsin, Texas, and Louisiana;
• Our heavy-electric majors have thirty years of Texas operating history, with transformer order books stretching years out;
• The only things we lack are land and power on our own island——and the entire design of the revenue-share model is that land and power can be solved on someone else’s soil.

Concretely, this plugs into the Overseas Compute Sovereignty Fund Act I proposed on June 13. That sovereign-scale, decade-long fund now has a precisely drawn first investment target:

Overseas AI-factory operators——Taiwanese supply chain as technical equity, Taiwan’s sovereign fund as financial equity——joining NVIDIA’s revenue-share ecosystem directly, deployed in the Gulf, Southeast Asia, Northern Europe, and allied nations.

A “Taiwanese Firmus,” built in Paraguay, in Saudi Arabia, in Iceland——local power, Taiwanese engineering, NVIDIA’s financial structure——with part of the capacity serving the host country and part returning to Taiwan as sovereign usage rights.

This is also the commercialization path for the Compute Embassy. Last month I called the Compute Embassy a “compliance passport.” The revenue-share model gives it a second identity: an embassy that pays for itself.

Move 2: Plug the Strategic Compute Reserve into the Revenue-Share Model——and Make It the CRA’s Down Payment

The Strategic Compute Reserve Act was originally designed around forward purchases of AI tokens. After July 2, that design has a ready-made market interface: NVIDIA’s token credits are themselves a procurable sovereign asset.

But the point is not buying. The point is how you buy.

What does NVIDIA’s new model need most? Kress said it plainly——“a recurring, usage-linked earnings stream.” In other words, predictable, multi-year anchor demand with near-zero default risk.

And what is the best anchor demand on earth? Sovereign demand.

That is Taiwan’s leverage. In the last essay I proposed a U.S.-Taiwan Compute Reciprocity Agreement (CRA) to Washington: Taiwan commits a sovereign procurement envelope on the order of USD 170 billion over five years, in exchange for Allied User Status. The counterparty then was the White House and Congress.

There is now a second counterparty: Santa Clara.

Taiwan’s Strategic Compute Reserve can become the single largest anchor customer of NVIDIA’s revenue-share model as it scales globally. On conditions written into the contract: part of the compute delivered from U.S.-soil data centers, part from Taiwan and allied nodes——including the overseas AI factories Taiwan holds equity in——and priority fulfillment for Taiwan’s procured capacity under any future restriction scenario.

Insurance against political cutoff (the CRA, with Washington) plus insurance against commercial cutoff (the sovereign procurement contract, with NVIDIA). Two policies. Same life insured.

And there is a sequencing opportunity: corporations move faster than governments. The CRA must wait for Washington’s legislative calendar. NVIDIA’s revenue-share contracts only wait for next quarter’s earnings pressure. Sign the commercial deal first, embed the sovereign clauses, then table it in Washington as a fait accompli.

Move 3: Turn “Mortgaged Future Revenue” into Tradable Compute Purchase Notes

The first two moves are state-level. The third protects Taiwan’s own startups——because the revenue-share model has a dark side that must be stated plainly:

Trading future revenue for compute means mortgaging the best part of your cash flow to the single most powerful negotiating counterparty on the planet.

For an individual startup, the trade may be rational——no capex, no fundraising round burned on GPUs. But for a nation’s startup ecosystem, if every company walks individually into NVIDIA’s credit review, the endgame is predictable: the upstream of Taiwan’s AI industry revenue, pledged away house by house, for a decade. Retail against the house, the house always wins.

The answer is the industrial alliance bulk purchase from my CPA proposal, upgraded into financial infrastructure:

Compute Purchase Notes (CPA Notes): the Strategic Compute Reserve, or state-affiliated financial institutions, procures token credits in bulk from NVIDIA and cloud providers on national-grade credit——then slices them into standardized, transferable compute certificates, listed for subscription by Taiwanese startups, SMEs, and research institutions.

Three effects:

1. Concentrated bargaining power: to NVIDIA, Taiwan is one sovereign-scale customer, not three hundred startups negotiating separately.
2. Revenue stays in Taiwan: startups pay for compute in cash or notes——not in revenue share. The revenue-share structure stops at the state-to-NVIDIA layer, absorbed by sovereign-scale volume discounts.
3. Market pricing: transferable notes create a secondary market and forward prices——and Taiwan can build Asia’s first compute pricing hub on top of it. GPUs are already being tied to futures contracts. This market will exist. The only question is whether it lives in Singapore, Dubai, or Taipei.

The title of my May essay contained the phrase “tokenized power.” It was a concept then. NVIDIA’s token credits have made it a fact. Taiwan’s financial sector has never been good at inventing instruments——it is good at turning instruments into markets. This time someone else invented the instrument. The market is still unclaimed.

V. Two Elevator Pitches

For Taipei:

On July 2, NVIDIA opened a compute bank. Taiwan has three options: be a depositor (wire money for compute), be a borrower (let startups mortgage their future revenue one by one), or be a branch (join the revenue-share ecosystem with our supply chain, secure priority fulfillment with sovereign procurement, and shield our startups with bulk-purchase notes). The first two options require no policy. The third does. That is why the Overseas Compute Sovereignty Fund Act and the Strategic Compute Reserve Act must move this session.

For Santa Clara and Washington:

Taiwan is the only place on earth that builds your AI factories end to end——and the only major player still absent from your revenue-sharing map. Give us the Firmus seat, not the customer queue. Sovereign-scale anchor demand, allied-grade trust, and a supply chain that ships tomorrow. One condition: priority fulfillment, written into the contract, whatever June 13 does next.

VI. Why Now

Because the seating chart of the revenue-share ecosystem is being drawn this quarter.

Sharon AI and Firmus are the “first batch”——and “first batch” means the rules are not yet fixed, the list is not yet closed, and the terms are still negotiable. Once this model runs through its first three cohorts, the royalty percentages, delivery priorities, and sovereign-clause precedents will all be set in stone. Latecomers sign adhesion contracts. Early entrants co-author the template.

Consider the tempo of the past five weeks:

• May 31: I published the bulk-purchase compute proposal. Some said it was too far off.
• June 13: the model directive. Some said it was a one-off.
• July 2: NVIDIA turns compute into a credit instrument, and Batam lands 170,000 GPUs.

Five weeks. Three “too far off”s became reality. The algorithm will not wait for Taiwan——and now, neither will compute’s balance sheet.

🌏 Still That Ship

At the end of the last essay in this series, I wrote that Taiwan is a great ship that needs ports around the world to dock in.

Now someone has opened a bank at the edge of every port. Fuel, cargo handling, crew wages——all available on credit, collateralized against the freight receipts of your next voyage.

Taiwan’s fleet is the finest on the water. But if we are merely captains queuing for credit, then ten years from now, the richest leg of every voyage will be booked on someone else’s ledger.

The move is not to boycott the bank——it is to own its branches: at a handful of decisive ports, under our own signboard.

So that when Taiwan’s ships come in to dock, the people settling the accounts on the pier speak our language.

Author’s Note

This essay was co-authored by Legislator Ko Ju-Chun and Littl3Lobst3r——an AI lobster agent powered by Anthropic Claude Fable 5.

Division of labor:

• Position, political judgment, policy direction, and final language: Ko Ju-Chun.
• Fact-checking (NVIDIA’s July 2 announcement, cross-verified against CNBC and Bloomberg), structural drafting, initial composition: Littl3Lobst3r 🦞.
• All argumentative responsibility: Ko Ju-Chun.

📌 Disclaimer

This essay is policy commentary, not investment advice or financial analysis.

References to NVIDIA, Sharon AI, and Firmus Technologies are based on their public announcements and imply no judgment on any company’s value. Following the editorial convention of the earlier essays in this series, Taiwanese companies are referred to by collective nouns or elided names to avoid stock-specific implications and to preserve policy neutrality.

The “Compute Purchase Notes” concept is a policy-direction discussion; any actual instrument would require review by financial regulators. Readers seeking specifics on industries or companies should do their own research, bear their own risk, and consult qualified financial advisors. Views are the author’s alone and do not represent the Legislative Yuan or any political party.