The GPU claim
TokenOS sold the newest NVIDIA chips to anyone with a wallet, in any quantity, in the year the largest buyers in the world said they could not get enough of them. It never said where they came from. And at the public on-demand rental prices, no GPU type could have earned what it paid out for it.
What a node was
A node was sold as one GPU, provisioned by TokenOS in a partner data center, working for paying compute customers for 30 days. The customer paid the price of the node up front, plus a small fee. Thirty days later the treasury sent back the price plus a return of about 27% to 40%, depending on the GPU type. The operator account said the hardware was leased, that the customer's money was the cost of the lease, and that it was paid out to the data center rather than held.
You don't need it. We provision the H100s in our partner datacenters. You simply rent and deploy them to the network.
We do not own the hardware, the $2500 is the hard cost of deploying leased GPUs in a data center.
It's not held. It's converted to USD and paid to the datacenter(s) that deploy the nodes.
A member had asked where the $2,500 was held and how safe it was.
Yea, as of now, our orchestration sees it as one GPU = one node
No, we do not have the hardware. Your initial thought was correct.
A member had asked whether the team had real hardware or rented and resold compute.
The prices and the payouts are TokenOS's own figures. The prices are in the public code of its site. The payouts are what the treasury paid per node before August 12, 2026, visible on the ledger.
| GPU type | Price of one node, 30 days | What the treasury paid back per node | The customer's return | Nodes sold |
|---|---|---|---|---|
| H100 | $2,500 | about $3,190 | about $690 | 414 |
| H200 | $3,125 | about $4,270 | about $1,140 | 1,187 |
| B200 | $5,250 | about $7,350 | about $2,100 | 5,147 |
| B300 | $7,500 | about $10,050 | about $2,550 | 200 |
| GB300 | $9,000 | about $12,600 | about $3,600 | 96 |
Where the chips were supposed to come from
TokenOS sold its nodes from December 2025 to September 2026. Three of every four were Blackwell chips, the B200, B300 and GB300, NVIDIA's newest. In that window two of the largest buyers of GPUs in the world, and the company that makes them, said on the record that there were not enough to go around.
- NVIDIA, November 19, 2025, announcing its quarterly results: "Blackwell sales are off the charts, and cloud GPUs are sold out."
- Microsoft, October 29, 2025, on its earnings call: "demand remains significantly ahead of the capacity we have available", and "we now expect to be capacity constrained through at least the end of our fiscal year."
- Alphabet, February 4, 2026, on its earnings call: "we've been supply-constrained, even as we've been ramping up our capacity", and "I do expect to go through the year in a supply-constrained way."
- Microsoft again, July 29, 2026: "Customer demand continues to exceed available capacity."
Those are companies with budgets in the hundreds of billions, buying directly from NVIDIA, saying they could not get enough. A single GPU, even a B200 or B300, could be rented by anyone on October 21, 2026; the price pages further down show that. The question is volume, and timing. The newest chips, in volume, went to the buyers NVIDIA served first, and even they were short.
In the same months TokenOS sold the newest chips to anyone with a wallet, "in 3 clicks", in any quantity a buyer would pay for, as if its supply had no limit. Single orders in its records run to 116 nodes. By July 2026 its customers were buying 2,741 nodes a month, 1,946 of them Blackwell. In all it sold 5,147 B200 node-months, 200 B300 and 96 GB300. It said there was a shortage, and used it as a sales pitch.
GPU shortage incoming? TokenOS.ai has you covered. Analysts predict a massive GPU squeeze in 2026. If you don't have access to compute, you're dead in the water. Secure your H100 allocation via TokenOS NOW. Don't get squeezed.
NVIDIA H100s are in short supply globally. TokenOS.ai gives you access to deploy this coveted hardware and monetize it instantly. Don't wait for backorders. Deploy now.
Stakers get priority access to B200 node allocations. In a world of hardware scarcity, access is the ultimate utility. Stake to secure your power.
Standard cloud providers require corporate contracts, credit checks, and non-disclosure agreements to access upper-tier Blackwell hardware. The upcoming DeAI Compute Marketplace destroys these barriers. If you have the wallet balance, you have the compute.
Where the chips came from was never said. In January 2026 the account said it used "5 different datacenters", which "could probably support up to 200 within 24hrs, but after 200 they may need to source more elsewhere." That was a rate for bringing nodes up, not a count of GPUs. By July 2026 about 2,700 nodes were supposed to be running at once. In June it said "we now have multiple different datacenters that we work with and its increasing weekly." In July, asked what would happen when supply ran short: "It's a very real problem we could be facing in 4-6mo or so, depending on the supply in the overall market." No data center was ever named. No contract with one was ever shown. No payment to one appears on the ledger. Members asked for a readout from the servers in August 2026; the account, writing as "engineering team member here", refused and called the request "a classic soft reconn pattern."
The chip that does not come alone
The GB300 is NVIDIA's newest Blackwell chip. NVIDIA sells it as a 72-GPU rack, the GB300 NVL72; the first such rack was brought up by the cloud provider CoreWeave in July 2025. TokenOS sold single "GB300 nodes" from February 25, 2026, at $9,000 each, 96 of them, and paid back $12,600 for each one that came due before the stop. On October 21, 2026, neither public price page offered a single GB300 to rent at any price, and Amazon listed none. In August 2026, after six months of those sales, the account explained the product:
There is no standalone GB300 product. You cannot buy a GB300 as a PCIe card or a single server. The Grace + Blackwell Ultra superchip ships only inside the NVL72 rack system, which is 72 GPUs.
Posted to explain why the GB300 had been removed from the menu.
What a GPU earns
Set supply aside and suppose every chip existed. GPU rental is a public market with published prices, so what a GPU can earn is checkable. Three price points are used here, all read on October 21, 2026. Two large independent GPU clouds, Lambda and Runpod, publish on-demand instance prices per GPU per hour; the table takes the higher of the two. The same two pages also list higher prices for committed multi-GPU plans and per-second serverless use, so the highest figure anywhere on either page is shown too. The third point is Amazon's published price for reserved GPU capacity, the most expensive of the three.
What the GPU had to earn depends on who kept what. Until spring 2026 the account said it passed everything through: "we paid 100% of the compute income to noderunners when we were in alpha testing phase". From May 2026 it described a split: "Net profit is split 25% to treasury, and 25% to $TOS stakers", and "Hardware node operators take 50% net profit." On the first description the GPU had to gross the payout. On the second it had to gross the price plus twice the return. Thirty days is 720 hours, so each gives an hourly rate the GPU had to earn every hour, with no idle time.
| GPU type | Needed per hour, 100% pass-through | Needed per hour, 50% split | On-demand instance, Lambda or Runpod | Highest price on either page | Amazon, reserved capacity |
|---|---|---|---|---|---|
| H100 | $4.43 | $5.39 | $4.29 | $6.16 | $5.97 |
| H200 | $5.93 | $7.51 | $5.29 | $6.35 | $7.90 |
| B200 | $10.21 | $13.12 | $7.99 | $9.86 | $14.21 |
| B300 | $13.96 | $17.50 | $8.99 | $10.65 | $16.15 |
| GB300 | $17.50 | $22.50 | not offered | not offered | not offered |
Read the table generously. It assumes the GPU is rented every hour of every day, that nothing is spent on the data center, power, network or the GPU itself, and that no customer ever goes elsewhere. Even so:
- At the on-demand instance price, no GPU type earns its payout. An H100 that never sleeps grosses about $3,090 against a payout of about $3,190; a B200 grosses about $5,750 against $7,350.
- At the highest price on either page, the H100 and H200 clear the payout. The B200 and B300, three of every four nodes sold, do not. Under the 50% split only the H100 clears, and only at Lambda's 16-GPU plan price. The H200, B200 and B300 do not.
- At Amazon's own prices the arithmetic closes for the H100, H200 and B200, by 5% to 11% under the split. It does not close for the B300. Those are the prices Amazon charges for Amazon's own cloud, hardware and name.
TokenOS was not selling Amazon's cloud. It was reselling leased bare metal, and it described its own price: "Elastic, scalable, and 40% cheaper than other solutions." It said idle hours were sold to spot markets, the cheapest hours there are: "When internal demand is low, we wholesale your compute to spot markets." At 40% below the on-demand instance price an H100 grosses about $1,850 a month, and a B200 about $3,450, against payouts of $3,190 and $7,350. The payouts needed a seller charging about what Amazon charges, and for the B300 more. TokenOS said it charged less than everyone.
The same arithmetic applies in reverse. A customer who wanted the compute an H100 node supposedly delivered could have rented the GPU directly, from the same public pages, for about $3,090 a month. TokenOS charged $2,500 for it and promised $3,190 back.
The answer TokenOS gave
A member ran this arithmetic in the group on May 10, 2026, using the same public price ranges. The account's reply was that confidential computing changed the numbers.
You're not factoring in TDX into these assumptions
Well, it's a bit different, because the indexes track bare metal industry wide averages and don't account for all the extras, things like TDX, specific protocols, etc.
TDX is a security feature of Intel processors that lets a customer's data be processed inside a sealed enclave. Neither Lambda nor Runpod lists a surcharge for it. No customer or contract for such a premium was ever posted in the group. TokenOS did publish a price list, on its own marketplace: about $5.54 an hour for an H100, $7.68 for an H200, $13.38 for a B200, $17.99 for a B300 and $20.81 for a GB300. Those prices would cover the payouts. On September 17, 2026, the same marketplace's own statistics page showed no GPUs online and no operators live. It listed 18 rentals in its lifetime, 30.7 GPU-hours in all. For the premium to close the gap at the on-demand instance price it would have to be about 3% on an H100 and 28% on a B200, every hour of every day, before any cost. Under the 50% split, about 26% and 64%.
The arbitrage story
TokenOS said the gap was closed by arbitrage. A system it called A²E, the "Arbitrage & Orchestration Engine," would buy compute low and sell it high. It would route each node to the best-paying work at every moment. In its own words:
It's an algorithmic trader for GPU power. It buys low and sells high. You provide the hardware. We provide the market.
Current average yield for H100 nodes is outperforming standard cloud rental by 15%. A²E arbitrage is working.
ROI right now is averaging 45.8%, or roughly $120/day per node.
Arbitrage in GPU compute is real, and it is small. The trade is buying committed capacity below the market and reselling it at the on-demand rate when demand is high. The spread is the whole profit, and the on-demand rate is the ceiling a walk-up customer pays.
TokenOS's version fails on its own numbers. A customer's $2,500 for 30 days is $3.47 an hour, about 80% of the on-demand instance price. That spread, captured every hour of the month, is about $590. The return promised on it was about $690, and from May 2026 the return was supposed to be half the profit. The promised $120 a day is $5 an hour, more than either page charges for an on-demand H100 instance. "Spot markets" are the cheapest hours in the market, not a floor. Wholesaling idle time there means selling at the lowest price there is.
"7,000+ nodes deployed"
The investor page leads with "7,000+ nodes deployed". TokenOS's purchase records list 7,045 nodes sold. 7,044 are single-GPU nodes, counting every 30-day purchase and every time a customer bought again with a payout. One is a full GB300 rack sold in August 2026 with "No expected ROI". The sales count in its records matches the number on the page. Nothing in the records, and nothing on the ledger, says how many GPUs there were.