ENGY: The $135 Million Bet on Half-Price AI

Engy is easier to use than Bittensor is to understand.

For a customer, Engy is a website and an API key. For the network, it is subnet 53, a market that tries to reward computers for serving correct AI answers.

Start with the familiar part. A developer sends a question to an API. Engy routes that request to a machine running the selected model. The machine returns an answer. Engy charges for the number of tokens read and written. Its API follows the OpenAI format, which makes switching relatively easy for teams that already use that standard.

Bittensor sits underneath that experience. Independent operators called miners can provide model inference. Validators are supposed to judge their work and assign weights. Those weights influence how subnet emissions are distributed. The product and the incentive system are connected, but they are not the same thing.

Hanlin AI presents Engy and TrajectoryRL as its two Bittensor projects. Public materials say the team includes engineers and researchers from major AI labs, but the site does not name the individuals or a contracting legal entity. The service terms point users to a single Engy contact address.

One identity warning matters. Before Engy, the same subnet number was known as EfficientFrontier, a trading-strategy project. Engy's public repository first declared the verified-inference project on subnet 53 on July 6, 2026. The slot was renamed and repurposed. Pre-July product descriptions, usage, and project claims belong to the former project, not Engy.

One API request crosses three separate tests.

The model must produce an answer, the network must test where that answer came from, and the incentive system must decide who deserves payment.

The first step is ordinary inference. Engy currently lists Kimi K3, GLM-5.2, and Qwen3.6-35B-A3B. Customers buy prepaid credits and pay by token. Purchased credits are refundable after usage is deducted, while promotional or granted credits are not. The terms provide no uptime guarantee.

The second step is proof. Qwen miners return a compact TOPLOC fingerprint alongside each completion. In plain English, the fingerprint is a trace of internal model activity. A verifier compares it with the expected model state. A matching proof makes it harder to substitute a cheaper or altered model while claiming to run the approved one.

The third step is scoring. A correct proof does not decide the value of the request. Engy's control plane still determines admission, routing, the billed amount attached to a successful request, and the score rate. The public specification says only successful requests with a positive billed cost are scored. If no billed traffic exists, weight falls back to the owner hotkey.

The distinction matters because “verified inference” can sound broader than it is. TOPLOC can help prove the model path. It does not independently prove that the customer was real, the price was collected, the score rate was fair, or the weight vector was calculated correctly.

Half-price AI is not just a slogan.

At the evidence cutoff, the posted Kimi K3 discount was exact. The more difficult question is who pays for it.

Engy charged $1.50 per million Kimi input tokens and $7.50 per million output tokens. Kimi's official API charged $3 and $15. Engy also listed GLM-5.2 at $0.68 input and $1.50 output, versus Z.AI's $1.40 and $4.40. Its Qwen rate of $0.045 input and $0.30 output was below Alibaba's international $0.375 and $2.25.

Those comparisons are real price differences. They are not proof that every service is identical. Context limits, quantization, latency, caching, routing, geography, uptime, support, and model version can change the experience. Engy's project-reported 90-day uptime was 96.98% for the gateway, GLM, and Qwen, and 92.68% for Kimi. There was no contractual service-level agreement.

How can Engy charge so much less? Start with the model itself. Kimi, GLM, and Qwen publish weights that can be run outside the model maker's API. Engy can therefore operate the inference rather than buy each answer from the company that trained the model. It avoids the model maker's API price and competes on its own hardware, power, utilization, and operating cost.

The project says its launch models ran on consumer-GPU clusters, including GLM on 32 RTX 5090s and Qwen on RTX 4090 hardware. It also uses lower-precision model formats. In plain English, smaller numerical formats let the same hardware hold more of the model and process more work, although they can also make the service different from the model maker's own deployment.

Bittensor adds a second possible source of payment. Engy sets a score rate for successful billed requests, validators turn those scores into weights, and the weights influence emissions. That means customer fees do not necessarily have to fund every miner, validator, and subnet-owner reward during the network's early stage.

The missing bridge is the size of each contribution. Public materials do not show model-level compute cost, paid versus promotional usage, customer acquisition cost, emissions received, direct subsidy, or gross margin. Kimi and GLM were also first-party lanes at the cutoff. Their discounts cannot yet be credited to open competition among miners. The honest conclusion is narrower: self-hosting and consumer hardware explain how a lower price is possible; public evidence does not yet prove the price is self-supporting.

Engy is being noticed by people who can send it customers and capital.

Attention is not revenue. But for a young infrastructure product, the right attention can accelerate trials, suppliers, financing, and token demand before those outcomes appear in financial data.

Jason Calacanis said he had tried Engy and described the combination of TAO and open models as a path toward lower token costs. This is a meaningful product-discovery signal. Calacanis is not simply a large account: his official biography describes him as a technology entrepreneur, fund manager, All-In co-host, and early investor in more than 350 startups. His post had attracted more than 90,000 views by the cutoff.

Mark Jeffrey, a partner at a Bittensor-focused fund, amplified the half-price comparison to more than 80,000 views. Algod, an influential trader with roughly 200,000 followers, argued that cheaper open-model tokens could increase total demand. These voices can spread the thesis. They do not independently verify Engy's cost structure, customer quality, or revenue.

The conversion test is simple. If attention is becoming adoption, Engy should show more paid accounts, repeat usage after credits expire, a broader model mix, and cash growth. If attention is reaching only the subnet market, alpha can reprice long before the business evidence improves.

Millions of requests can still be a small business.

AI usage is measured in tokens, not request headlines. The cheapest model generated almost all requests, while the premium models generated much of the implied dollar value.

Engy's provider dashboard reported 2,265,520 requests over the 48 hours ending at the cutoff. It showed a 92.5% success rate, 28.0 billion prompt tokens, 1.19 billion completion tokens, and 1.61 million project-reported audits with a 100% pass rate.

Qwen accounted for 2.24 million requests, or 99.0% of the total. GLM contributed 15,284 and Kimi 8,375. That concentration is not automatically bad. A cheap, useful model can be the wedge. It does mean the headline request count says more about Qwen traffic than broad product adoption.

The dashboard does not disclose unique users, paying accounts, requests per customer, promotional credits, retention, cohort growth, cash receipts, or acquisition cost. It also does not publish third-party audit results. “Live paying buyers,” as claimed in the project announcement, is credible first-party evidence of commercialization, but not a customer or revenue audit.

We translated the reported prompt and completion tokens through Engy's posted list prices. Qwen produced about $1,557 of list-price-equivalent usage, GLM about $326, and Kimi about $1,325. The total was roughly $3,207 for 48 hours.

If that exact window repeated for a year, it would equal about $585,000. It is not a forecast and it is not reported revenue. The calculation does not subtract cached-token discounts, free credits, failures, refunds, taxes, price changes, or uncollected balances. It also does not add any off-dashboard contracts. It is simply the best reproducible bridge from public usage to a possible billing ceiling.

The network verifies a signature more independently than it verifies a score.

Engy's public code is unusually candid about its current centralization. That candor is a strength, but it also defines the work still ahead.

The repository is public and includes the miner client, validator, model specifications, tests, and a one-page protocol description. No license file was present at the reviewed commit, so the code is publicly inspectable but should not yet be described as fully open source.

TOPLOC is enforced now for Qwen. A second verifier based on sampled GPU matrix multiplication is described as rolling out. The README says the audit verifier, full miner client, and full incentive mechanism will be published later. Kimi and GLM had no public audit counts at the cutoff.

The light validator downloads an epoch result, checks that it was signed by a pinned master hotkey, and submits the weights. The code is explicit: this is not an independent rescore. It can detect the wrong signer or a forked result. It cannot detect a master that scored traffic incorrectly.

Epoch 7 was finalized with 144 miners, 2.60 million requests, and a master signature. Epoch 8 was open and unsigned at the cutoff, with 91 miners and 48,219 requests. The protocol is therefore distributed at the compute edge, but concentrated at admission, routing, billing, scoring, and weight formation.

Buying SN53 alpha is a swap into a subnet market, not a share purchase.

The easiest mistake is to treat the subnet's alpha token as stock in Engy. It is not documented that way.

Dynamic TAO gives each subnet its own automated market maker. One side holds TAO. The other holds that subnet's alpha. The reserve ratio sets the alpha price, and staking TAO into a subnet buys alpha from the pool. Unstaking sells it back.

SN53's alpha symbol is ب. At the cutoff, one alpha cost 0.032763 TAO. TAO.app identified subnet 53 with owner coldkey 5H8SFV…VL8Q and hotkey 5DXSBC…1uvJ. Identity matters because alpha exists by subnet number, not as a standard exchange ticker.

How to buy: first buy TAO on an exchange that supports withdrawals. Move it to a Bittensor-compatible self-custody wallet. Open TAO.app's subnet 53 page, connect the wallet, choose Buy, enter the TAO amount, review the alpha received and price impact, then confirm. Centralized exchanges generally sell TAO, not SN53 alpha.

The transaction behaves like a swap. The SDK applies an alpha transaction fee of about 0.05% when applicable, and the pool creates slippage. A large buy lifts the price paid. A large sale pushes it down. Selling requires reversing the trade through the same market.

What does alpha own? Public documents do not grant equity in Hanlin AI or Engy, a share of API revenue, customer credits, governance rights, intellectual property, or an enforceable claim on cash. Alpha is exposure to a subnet market and its emissions. That can still be valuable, but it is a different asset from company ownership.

The subnet is worth $7 million, $36 million, or $135 million, depending on the question.

All three numbers are valid. None should be presented without its label.

Pool liquidity: the subnet held about 35,340 TAO and 1.08 million alpha in its reserves. At $196.93 per TAO, the TAO side was worth roughly $7.0 million. This is the most relevant number for trading depth and exit capacity.

Issued-supply value: TAO.app showed 5.64 million alpha issued, about 26.8% of the 21 million maximum. Applying the spot exchange rate produced 184,640 TAO, or about $36.4 million. This is closest to a circulating or issued market-value lens.

Fully diluted value: applying the same spot price to all 21 million alpha produced 688,020 TAO, or about $135.5 million. This is the number in the title. It assumes the entire future supply can be valued at today's marginal pool price, an assumption that becomes less realistic as size grows.

The gap is the thesis. A roughly $135.5 million FDV sits beside a reproducible public usage window that annualizes to about $585,000 of list-price-equivalent billings. Dividing one by the other would create a number near 232 times, but it would not be an equity revenue multiple. Alpha holders do not own that billing stream, and the billing estimate is not audited revenue.

The correct use of the comparison is not “cheap” or “expensive” by itself. It shows how much future proof is embedded in the price: more real customers, durable retention, sustainable discount economics, broader miner participation, independent scoring, and a clearer connection between network value and alpha demand.

The bull case is a cost flywheel. The bear case is a subsidized funnel.

Bull case: Engy makes a hard product feel ordinary. Developers get familiar tooling and better prices. Consumer GPUs lower compute cost. Bittensor emissions recruit more providers. More supply improves availability. More customer volume improves utilization. The resulting cost advantage attracts still more demand.

If that loop works, Engy does not need to beat every frontier lab. It can become the low-cost distribution layer for strong open-weight models. The team's other subnet, TrajectoryRL, adds a possible research advantage. Model training, inference, and verification could reinforce one another.

Bear case: the discount is funded by emissions or early subsidy rather than structural efficiency. Qwen traffic is concentrated, low-value, or promotional. Premium models remain first-party. Customer retention is weak once centralized labs change prices. The status record and lack of an SLA limit production adoption. Weight formation stays concentrated at the master.

The bear case does not require the API to stop working. A useful, growing product can still be a poor alpha investment if users do not need the token, cash flow does not reach holders, and future emissions dilute demand faster than the network creates it.

Six facts will decide whether the $135 million bet earns its proof.

The next update should not ask only whether requests went up. It should test who paid, who computed, who verified, and what alpha holders actually gained.