The Paris 2026 programme ran to 77 sessions over two days. Sixteen of them carried the Bittensor track, more than the twelve that carried stablecoins, and eight ran on a stage named the Taostats Stage, after a company few institutional readers will know. The same programme carried speakers from the Bank of England, JP Morgan and BlackRock. That combination deserves an explanation.
Proof of Talk is an application-only summit staged at the Musée des Arts Décoratifs in the Louvre Palace in Paris and, from December, at the Louvre Abu Dhabi. Capacity is capped, admission is by application and review, and no speaking slot is sold, so a track this size is a fact about editorial judgement, not about who paid. Our Bittensor track page lists who spoke and who partnered. It does not explain the protocol. This does.
What was actually on the programme
The track opened at 13:00 on Day 1 with a machine-learning workshop in the Roundtable Lounge, and five minutes later Ala Shaabana, billed as a co-founder of Bittensor, took the main stage for a fifteen-minute keynote, Subnets, Signals, and Scale: The Crucible of Open Intelligence. His co-founder Jacob Steeves had the same stage on Day 2 at 10:05 for a thirty-minute fireside, Decentralising Intelligence: The Unification of AI and Bitcoin, moderated by Mark Jeffrey of Stillcore Capital. Between those two main-stage appearances sat a seventy-five-minute block of subnet keynotes from six teams, the panel Subnets as an Asset Class: Building the On-Ramps for Institutional Capital, a confidential-computing workshop run by Targon and a live pitch competition, all on the agenda and the session pages.
One detail stands out. The Thin Line, Marcus (Mog) Graichen's Day 2 main-stage keynote, is the only session in the 77-row programme carrying a written description: "Why the agent economy needs both a permissionless protocol and an accountable industry built on top of it, and why Bittensor is the only network architected for both." It is a claim from a stage, not a finding, and Graichen was billed as founder and chief executive of Taostats, a track sponsor and a primary source of the ecosystem's data.
What Bittensor claims to be
The network's own account of itself is a workshop paper, Incentivizing Intelligence: The Bittensor Approach, written by Steeves, Shaabana and four Opentensor Foundation colleagues and presented at TSRML 2022, co-located with NeurIPS 2022. It opens with an analogy: "Inspired by the efficiency of financial markets, we propose that a market system can be used to effectively produce machine intelligence."
Strip the vocabulary and the proposition is straightforward. Machine intelligence is expensive to produce and hard to evaluate. Closed laboratories solve the evaluation problem internally, paying people to judge output. Bittensor proposes an open market instead: anyone may supply a model, other participants grade it, and the network emits a token to whoever grades well. The network pays, continuously, in its own unit, for output it has scored.
How the incentive works
The clearest plain-English description is not in the project's own material but in Grayscale's Form S-1/A, filed with the SEC on 2 April 2026: "Miners host AI models and make them available to the network. Miners form coalitions (known as Subnets) based on the types of tasks they seek to perform." A subnet is a task-specific market.
The official documentation supplies the arithmetic. TAO has a 21 million maximum supply, as does each subnet's own alpha token. A block is produced every twelve seconds. Base emission was one TAO per block; the first halving occurred in December 2025, taking it to 0.5 TAO per block, roughly 3,600 TAO a day. Within a subnet, per epoch, alpha splits 41 per cent to miners, 41 per cent to validators and their stakers, and 18 per cent to the subnet owner.
Grading runs on Yuma Consensus. As Taostats documents it, in a page last updated on 20 August 2026, "validators test the output of miners, and grade the results as weights", validators earn dividends for evaluations "in agreement with the subjective evaluations produced by other subnet validators, weighted by stake", and weights out of consensus are "given less 'weight'". Taostats is Graichen's company and a track sponsor, so read that as the ecosystem describing itself, not as independent verification. Since 14 February 2025 each subnet has also had its own token and liquidity pool. As the explainer Tokenomist put it on 2 April 2026, "No committee decides which subnets get funded. Your stake does."
Pricing machine intelligence means two things. Inside a subnet, validators price a miner's output. Across subnets, a token market prices the subnet itself, and that price sets its share of the daily emission. The first is a quality signal; the second is a capital-allocation signal only as good as the market making it. Tokenomist, sympathetic to the design, concedes that the system "cannot fully distinguish whether that flow is driven by real demand or coordinated activity", and that "most subnets today are still pre-revenue".
What the independent evidence shows
The most substantial empirical work is independent. Bittensor Protocol: The Bitcoin in Decentralized Artificial Intelligence? A Critical and Empirical Analysis, by Elizabeth Lui and Jiahao Sun, submitted on 29 June 2025, parsed 6,664,830 on-chain events from 20 March 2023 to 12 February 2025 across all 64 then-active subnets and 121,567 wallets. Its abstract: "we first document considerable concentration in both stake and rewards. We further show that rewards are overwhelmingly driven by stake, highlighting a clear misalignment between quality and compensation."
The top 1 per cent of wallets controlled between 38.5 per cent and almost all of the stake in a subnet, with a median of 89.8 per cent, at a mean stake Gini coefficient of 0.9825. For miners, the correlation between stake and reward ran 0.50–0.80, between measured performance and reward only 0.10–0.30, and between stake and performance approximately zero. Over half of subnets required under 1 per cent of wallets to amass 51 per cent of stake.
Two caveats, both cutting in the network's favour and both usually omitted. The data window closes on 12 February 2025, two days before the per-subnet token market went live, so it measures the previous regime rather than the system the Paris speakers were arguing about. And both authors are credited to FLock.io, which builds in the same category. It remains the best independent measurement we found, and a taxonomy paper by an overlapping group, SoK: Blockchain-Based Decentralized AI, devotes an appendix to Bittensor's security concerns, with sections on centralisation and collusion.
The sharpest critique, though, is in a filing by a party seeking to list the asset. Grayscale's own risk factors warn that dTAO "may concentrate influence among a small number of large TAO-holders", that emissions "could become concentrated in speculative or underperforming subnets", and that Yuma Consensus, proof of authority and dTAO "are new blockchain technologies that are not widely used".
The revenue question, and the measurement problem
On 23 March 2026, ten weeks before the Paris edition, Pine Analytics published a bear case estimating whole-network external revenue at 3 to 15 million dollars a year against an annual incentive budget of roughly 360 million. Chutes, a serverless inference subnet, took 14.4 per cent of emissions, about 142,000 dollars a day, against estimated external revenue of 1.3 to 2.4 million. In its words, "for every dollar customers pay, the network contributes $22–40 in TAO emissions." On defensibility it is blunter: "Bittensor subnets build none. The models are open source, the APIs are standard, and users can migrate to any provider serving the same weights with zero friction."
Those comparisons rest on estimates, as Pine Analytics says plainly: "There is no aggregated dashboard tracking external revenue by subnet. AI service delivery (inference requests, compute jobs, training calls) happens off-chain and is not recorded on the blockchain." On its assessment, widely quoted subnet revenue figures are projections repeated across analyst reports, not audited results.
The opacity is not confined to revenue. Grayscale's April filing put circulating supply at approximately 11 million TAO as of 31 December 2025; CoinGecko showed 9.597 million on 5 September 2026, a lower figure nine months later for a supply that only grows. TAO Media acknowledged the conflict on 22 August 2026, citing provider estimates from 9.6 to 11.3 million.
Governance, and what happened in 2026
Less than two months before the Paris edition the governance question became concrete. Yellow.com reported on 10 April 2026 that Covenant AI, the group behind the Covenant-72B model, had left the network, stating that emissions to its subnet were suspended, its permissions revoked and protocol changes made without its input, and calling that centralised governance. That is a statement by a party with a position, and the report documents no response. The underlying question is the one institutions ask first: who can switch off a participant, and by what process.
The question surfaced from inside the room too. Yuma, the Bittensor-focused subsidiary Digital Currency Group launched in November 2024, opposed the Root Reborn upgrade on 18 June 2026, warning that "moral hazard is acute" where validators can direct allocations toward subnets in which they already hold positions. Crypto.news, reporting it next day, called Yuma the network's third-largest validator. Evan Malanga, billed as Yuma's chief revenue officer, had sat on the Subnets as an Asset Class panel sixteen days earlier. Security history matters: The Block reported on 4 July 2024 that a malicious PyPI package posing as a legitimate Bittensor release drained about 8 million dollars in TAO from users before the chain went into safe mode.
What is demonstrably running
Some subnets ship verifiable product, and two were on the Paris stage. Score is subnet 44, doing computer vision on football footage; its open repository documents the pipeline and a January 2025 mainnet deployment, and The Block's coverage of Yuma's State of Bittensor report on 5 March 2026 reported that Reading Football Club had tapped Score's video product. Both Score founders were on the programme, Tim Kalic in the Day 1 subnet keynotes and Max Sebti on the Day 2 enterprise panel. BitMind is subnet 34, and a paper submitted on 14 July 2026 by its founder Ken Miyachi and a colleague reports its detector matching the best commercial deepfake detector tested on images, 0.915 AUC against 0.90. That is BitMind measuring BitMind. The same March report recorded de-concentration, the top 20 per cent of subnets capturing over 55 per cent of aggregate subnet value against over 82 per cent earlier, alongside The Block's caution that adoption is "arguably slow-going compared to the rapid uptake of closed-source AI models like ChatGPT". Subnets also disappear, into a live deregistration list.
Why an institutional programme carried it
Three interests were in the room and should not be conflated. The first is investment. Grayscale's April 2026 filing seeks to move its Bittensor trust from OTCQX to NYSE Arca. TAO Synergies, listed on Nasdaq, reported digital assets of 16,809,405 dollars and staking revenue of 477,400 dollars for the six months to 30 June in its 10-Q filed on 13 August 2026. Real numbers, and small ones.
The second is infrastructure. Confidential compute, GPU marketplaces and distributed training are procurement questions, not thesis questions, and they are why a Big Four partner turned up. Enterprise Adoption of Bittensor put Jean-Thomas Ledore of PwC and Shane Smith of Alera Group on a panel with subnet operators, moderated by James Altucher of TAO Synergies. That is early diligence, not deployment, and it is the half of the programme our roster by institution leaves out.
The third is curiosity, a reason most conferences will not admit to. Ten weeks after the Pine Analytics bear case, Sami Kassab, billed as managing partner of Unsupervised Capital, gave a ten-minute keynote arguing the bear case was stale. We are not claiming he was answering that piece. Only that the argument was live enough to schedule, and that his firm publishes its own investment case for the same network.
The sponsorship question, asked directly
A reader should ask whether this track existed because it was paid for. A stage carried a sponsor's name, a workshop was branded to a subnet operator, and the pitch competition ran with a crowdfunding platform. Our policy is that speaking slots are not for sale at any price, set out in how the programme is built. It governs who speaks, not what a stage is called or who hosts a side event. On this track more than most, check the affiliation of anyone making a claim, including in the sources above.
Where that leaves it
Published coverage, catalogued in what the record shows, captured the shape of these sessions and almost none of their substance.
The sober position in September 2026 is neither dismissal nor endorsement. Bittensor is a running system with a coherent mechanism, measurable concentration, an unaudited revenue base, contested governance and a handful of subnets with real customers. It belonged alongside a central-bank supervisor and a global custody bank for precisely that reason: it is the clearest current attempt to build a market for something that has so far only been bought under contract. The next edition is at the Louvre Abu Dhabi on 3–4 December 2026, by application and review.