Inference

Inference, or Inference.net, is the company that launched Kuzco, a decentralized GPU network on Solana. This is the AI compute project, not the Disney character or the lighting brand. Kuzco went live on 6 March 2024. People connected their GPUs to it, ran open-source language models such as Llama 3 and earned KZO points. Later that network ran as the "Inference devnet", with test staking on Solana Devnet. By 2026 the company had moved to a centralised business: it trains and hosts custom AI models for companies. As of October 2026, the devnet website and its documentation both redirect to the company's X account, and no token has launched. Treat the Solana network as inactive.

What happened to Kuzco?

Kuzco was started by Sam Hogan. In a November 2024 letter to the community, he wrote that the network grew to "over 8,000 unique devices at its peak", had produced over one trillion Llama 3 8B tokens, and had also "had some major outages". Epoch One ended on 18 November 2024. Epoch Two was meant to serve real paying demand: the plan was that operators with stable throughput could join a "Kuzco Mainnet" and earn USDC as well as points.

The project then became Inference.net, and the Kuzco X account became @inference_net. The devnet documentation dates Devnet Epoch 3 to 6 June 2025. In November 2025, when the company announced its LOGIC verification method, it said it had run "a globally decentralized AI inference network with up to 10,000 heterogeneous GPUs" over the previous 18 months (blog).

After that, the company's public messaging moved away from the network. Neither its seed announcement (October 2025) nor any blog post since mentions Kuzco, Solana or a token.

How the Kuzco GPU network worked

• Workers: operators installed the Kuzco (later Inference) worker on NVIDIA or AMD GPUs or Apple Silicon Macs (the devnet docs set a minimum of 16GB of VRAM) and served LLM requests that the network sent to them.
• Points: in the Kuzco era, workers earned KZO points based on the work they did. Hogan's Epoch Two plan offered more KZO points for high-end cards such as A100s and H100s.
• Epoch 3 staking (Solana Devnet only): according to the devnet docs, each operator ran a staking pool. Delegators could stake $INT-DEV test tokens into these pools. Jobs went to operators based on device VRAM, total stake and reputation, and operators could share their USDC earnings with delegators. The docs said the test tokens "have no monetary value and should not be used for real-world transactions or bought or sold by anyone."

Is the Inference devnet still running?

Probably not. As of 3 October 2026, devnet.inference.net and docs.devnet.inference.net both return a permanent (301) redirect to x.com/inference_net. We could not find a single devnet, epoch or operator post on the company's X account in the 12 months to October 2026. We also found no shutdown announcement, so we can't give a closing date. If you still run a worker, assume it earns nothing until the company says otherwise.

Kuzco airdrop, KZO points and the $INT token

The KZO points earned in the Kuzco era, and the "$INT points" later shown on the devnet site, were never converted into a tradeable token. As of October 2026, none of Inference's official channels has announced a token generation event (TGE) for KZO or $INT. Treat any token, presale or "claim" site that uses the Kuzco or Inference name as unofficial unless @inference_net announces it.

What Inference.net offers now

Inference.net now sells AI infrastructure to companies that build AI products. It is a normal software business and doesn't run on Solana:

• Custom model training: the company trains small, task-specific language models on a customer's own production data, to replace expensive general-purpose models. In April 2026 it launched Catalyst, a platform for monitoring, training and deploying these models (announcement). At launch, the company said training would cost from about $25 to several thousand dollars, and that it would cover training and deployment costs during the public beta (blog).
• Gateway, tracing and evaluations: customers send their AI API traffic through Inference's gateway, which records each call so it can be used to evaluate and train models. AutoEvals, launched in August 2026, compares candidate models on a customer's live traffic (docs).
• Dedicated deployments: hosting for open-weight and custom models on dedicated GPUs.
• Schematron: the company's own open-source models that turn web pages (HTML) into structured JSON data. They were first released in September 2025, and Schematron V2 followed in April 2026 (blog).

The pricing page lists a free pay-as-you-go tier, a $250-a-month Growth plan and custom Enterprise contracts. It doesn't mention paying in crypto.

Inference.net funding, investors and founder

Inference raised an $11.8 million seed round, announced on 14 October 2025. Multicoin Capital and a16z CSX led it. Topology Ventures, Founders, Inc. and a group of angel investors also took part (Phemex News). The announcement says the money will go into custom model training and deployment, not the GPU network.

Sam Hogan is CEO. In October 2025 the company's X account listed its core team as the CEO, a CTO and three founding engineers covering protocol, machine learning and infrastructure. Its code is published under the context-labs GitHub organisation.

Who Inference.net is for

• AI product teams with high, repetitive model usage (pulling data out of documents, classifying content, running agents) who want to replace expensive general-purpose model calls with a cheaper custom model. This is what the company sells today.
• Not for GPU owners looking for crypto income. The worker network has gone quiet and no token exists. If you want to rent out a GPU on Solana today, look at the alternatives below.
• Not a Solana integration. As of October 2026, none of the current products uses Solana.

Inference.net risks and limitations

• The network looks abandoned, with no formal ending. The devnet sites now redirect to X, and we found no sunset notice. Points holders have no published conversion or compensation plan.
• Token scams. Kuzco ran a popular points programme and never launched a token, so fake KZO or $INT tokens and "airdrop claim" pages are likely. Only trust an announcement from @inference_net.
• No mainnet programs. Epoch 3 staking only ever ran on Solana Devnet with test tokens. We know of no mainnet program for it, and none is listed for this project.
• Reliability record. In its own November 2024 letter, the team admitted to "major outages" during Epoch One.
• Centralised service. The current platform is a standard hosted service. Customers trust the company with their production data and model weights. Read its terms and data-retention settings before routing sensitive traffic through the gateway.

Inference.net alternatives

• Solana GPU networks you can still contribute to: Nosana pays GPU hosts for AI inference jobs on Solana. io.net brings together GPU clusters for AI workloads. Both have live tokens and their own incentive schemes. Check what each pays today and what hardware it needs before you commit.
• Hosted inference and fine-tuning: for teams, Inference.net competes with other centralised providers of open-model inference and fine-tuning. It also competes with the frontier-model APIs it aims to replace. Its pitch is specialised small models that cost less than calling a general-purpose model.