Cleared buybacks, stocks as collateral, and agents as the customer: September 20-27, 2026

From September 20, 2026 at 6:00 p.m. through September 27, 2026 at 6:00 p.m. in UTC-05:00, the staff of the U.S. Securities and Exchange Commission answered two questions that had been sitting on top of DeFi's revenue models, and Aave began accepting tokenized American stocks as collateral for loans on Base. The first said that once a crypto system is functional, announcing a token buyback is not a promise of managerial effort; the second is the first live market where a share of Apple or Nvidia can be pledged to borrow dollars without being sold. 12 Four of the seven people tracked here read those two events as one story: the assets and the tokens that represent them are being wired into the same rails, and the legal question marks attached to each are being erased one staff document at a time. The other three argued about who ends up on those rails — software agents, retail investors, or nobody at all if the money stops printing.

Brian Armstrong, co-founder and CEO of Coinbase, spent the week describing a customer that does not fill in a form. On September 25 he amplified his own company's numbers on which software agents use to trade: "Grok is the leading client for agentic traders on Coinbase currently." The post he quoted put Grok's share of agentic trading on Coinbase for Agents at 59.54%, ahead of custom command-line tools at 21.73%, Claude at 13.37%, Perplexity at 3.48%, ChatGPT at 1.54% and Claude Code at 0.35%. 318 An agentic trading client is the interface an autonomous program uses to place orders; the measurement says more about which model has the best trading tooling than about who is ultimately taking the other side of the trade.

The same day he took a market call from an equity analyst and extended it. Quoting a post that BlackRock's research had named AI agents as a potential major source of stablecoin demand, Armstrong wrote: "1) The number of agents will keep growing 2) The number of agents that need to transact then keeps growing 3) Crypto and stablecoins will be their go-to payment rail. Yep." 419 That is a thesis with a checkable step: an agent that pays in stablecoins shows up as onchain transfer volume, and a payment rail chosen by machines looks different from one chosen by people, because machines pay in small, frequent, precisely-timed amounts.

Two posts on September 27 turned the observation into a product argument. "Every backend service should have a /feedback endpoint," Armstrong wrote. "Agents are quickly becoming the heaviest users of most APIs. When one hits a missing feature or a bug, it should be able to say so right there, in a structured way. If the request makes sense, another agent drafts the PR and a human approves it." On September 26 he pointed readers at the IPO feature inside the Coinbase app, which the company built so that retail users can take allocations that used to go to institutions. 20 He followed it with a second: internal platforms should publish uptime, latency and cost, take feature requests in the open, and be allowed to charge internal teams, because "agents can discover services, compare them, and route around the bad ones." 2122

On the asset side, nothing in the window was new except the volume. Armstrong answered a thread on Coinbase's tokenized stocks — which a Base contributor described as "real equity securities backed 1:1 by US stocks held in regulated custody, not synthetic trackers," with dividends and corporate actions passed through on Base — with a single word: "Yep." 2324 On September 25 he pointed at Uniswap's figure that tokenized equities on Base had crossed $300 million in volume. 525 On September 26 he congratulated Aero on a billion dollars of tokenized-stock volume in its first month, with the line that carries his week's actual claim: "Equity markets are moving onchain." 2627 A one-month-old venue reaching a billion dollars in notional is the kind of number that is easy to measure and easy to over-read, because tokenized-stock volume counts both directions of a trade that may be arbitrage between the token and the share.

The check on the agent thesis is whether it shows up in money. If agents are the heaviest users of Coinbase's APIs and its trading surfaces, that should appear in the company's reported volumes and in the mix of order sizes, not only in a percentage split of client software. The check on the asset thesis is the same one that has been open since August: the single-stock perpetuals Coinbase filed for still need the Commodity Futures Trading Commission to let them list.

Hayden Adams, who invented the Uniswap protocol and runs Uniswap Labs, gave the week's sharpest reading of the SEC staff's September 25 document in one line: "Some bangers from the SEC today: 1) buybacks do not make a commodity token into a security 2) liquid staking tokens for commodities are not securities." 6

Both of those are real, and both come with the caveat that the document states about itself. The answers "represent the views of the staff of the Division of Corporation Finance," are "not a rule, regulation or statement of the Securities and Exchange Commission," and "have no legal force or effect." 1 The first answer Adams cited appears in the FAQ on buybacks: where a crypto system is functional, "an issuer's announcement of a non-security crypto asset buyback program would not constitute a representation or promise to undertake essential managerial efforts." Where the system is not yet functional, the same announcement could count if the issuer presents the buyback as creating yield or return. The second sits in the section on classification: a staking receipt token can be a "digital tool" evidencing ownership, or a digital commodity when a protocol-based liquid staking provider issues it, because it derives value from a crypto system that is functional. 1

Why read a staff FAQ closely? Because it is the current answer the market prices against, and because it may not be the durable one. Commissioner Hester Peirce, whose statement on the innovation exemption Adams quoted approvingly the previous week, filed her resignation the same day; her last day is October 2, which will leave the commission with two members. 28 A staff view is not a rule, and the person most associated with the views the industry has been relying on is leaving.

Adams spent the rest of the week on scale claims. On September 25 he quoted an ARK Invest analyst's comparison with one word, "Not bad": Uniswap settled 140 million trades in August, "more than Cboe BZX and NYSE American combined," against roughly 2 million a month five years earlier, with NYSE Arca at 232 million as the next rung up. 729 On September 22 he responded to a Token Terminal post that Uniswap holds more than 99% of Robinhood stock-token deposits in DeFi on Robinhood Chain — $53 million in its v4 pools, $25 million in v3, with total deposits up 532.8% over 30 days — with "chat is 99% market share high?" 3031

A trade count is not a volume and a deposit is not a fee. Trades and deposits both flatter a venue that is being used to move assets between a chain and a brokerage, so the figures worth following are the ones a third party can reproduce: how much fee revenue the venue keeps, and how much of the tokenized-stock activity is round-tripping rather than new exposure. Both the 140 million figure and the 99% figure came from analysts or data providers with their own methodology; neither has been independently reproduced in the window.

Stani Kulechov, founder and CEO of Aave, spent the week on the asset side of the same story. On September 25, a dedicated Aave V4 market on Base began accepting seven Coinbase-issued tokenized stocks — Apple, Amazon, Alphabet, Meta, Microsoft, Nvidia and Tesla — as collateral, letting eligible non-U.S. users borrow USDC against them. 232

Kulechov's framing is the part worth keeping. "The prevailing thesis is that onchain finance will transform brokerage, particularly lending and borrowing," he wrote. "Aave has processed around $4T in deposits and $1T in borrowing, powered simply by onchain code. Onchain rails enable faster, more accessible ways to trade, lend, and borrow by reducing reliance on intermediaries. Now its time to expand these benefits to stocks. Over time, this has the potential to fundamentally reshape the cost structure of securities finance, creating more open, efficient, and deeper capital markets." 2 He also flagged why the SEC staff's buyback answer mattered to Aave specifically: "Huge win for DeFi. Especially removing the uncertainty around buybacks and maintenance work around the protocols." 32

The metric he says he watches is one he invented: "I like to measure them through addressable collateral. The bigger the universe of assets that can be used as collateral, the bigger the lending opportunity." He lists the progression as crypto first, then securities through Coinbase Tokenized Stocks and Horizon's real-world assets, and eventually solar, batteries, GPUs, robotics and space infrastructure, with Aave aiming to "accelerate it by a decade". 8 An addressable-collateral metric measures the size of the prize, not whether any of it borrows, so it is a ceiling rather than a forecast.

The protocol's own numbers moved. On September 26 Aave V4's public launch was described by an Aave contributor as six months old with deposits at $1.3 billion, up from the $1 billion the market crossed a week earlier; Kulechov's reply noted that the code had been published as feature-complete in summer 2025 and that "after few months since public launch the protocol has been growing well." 3334 On September 27 he added the number that matters more for the thesis: "Around half of the V4 users are net new Aave users meaning that the protocol has expanded the user reach for Aave." 9 Deposits that come from new users are a different thing from deposits that migrate from Aave's older markets, and the two would show up differently in total protocol revenue.

His one fight was about definitions rather than markets. Responding to a classification of vaults, Kulechov argued that a vault where a curator has discretion over allocation and can expand beyond a user's mandate should not be treated as non-custodial merely because it has a timelock: "Vaults that could reasonably be considered non-custodial are those without a manager." He closed by saying industry standards are worth developing "in a way that serves the broader industry rather than your own interests." 35 The dispute is not cosmetic. If discretionary vaults are treated as non-custodial by default, the assets in them sit outside the disclosure and liability regime that the rest of the market is being fitted to.

Changpeng Zhao, who co-founded Binance, gave the window's one long-form interview that can be sourced end to end. When Shift Happens, hosted by Kevin Follonier, published a 64-minute episode with Zhao on September 25, with a published chapter list covering the Binance founding story, the FTX episode, his prison term, the pardon, Hyperliquid's rise, why HYPE is still not on Binance spot, and what happened on October 10, 2025. 1036 Zhao's own post the next day was characteristically flat: "Full 64 minute interview. I haven't watched it yet. Punchy sizzle reel up front, but we did cover a lot of ground." 37

Two passages in the conversation are about the part of an exchange's business that affects investors most: which tokens get listed, and why so many of them go badly. Zhao's answer was that listings are a sample of a base rate. "If you look at any market, any industry, if you look at by project numbers, the majority projects will fail," he said. "If you look at internet, how many internet companies were there? Probably millions. And then how many are successful today? Probably like a few dozen. Well, a few hundred, maybe." The difference with tokens, he agreed when the host raised it, is that the losers are public and anyone can buy them. 10

From there he argued against the gatekeeping: "I actually believe that accredited investor protection, like to prevent you from investing early, is net going to make you poor." His conclusion was that access plus education beats exclusion — "give people access, give people education and let people have the opportunity to decide for themselves." 10 Asked whether retail can win against market makers, key-opinion leaders, launchpads and exchanges, he said yes in the long run and that the market self-selects, because over-hyped projects do not last.

He also gave a number for the education project he calls his highest-impact long-term work, Giggle Academy: about 100,000 children at the start of 2026, and 1.3 million now. His critique of schooling was that it aims at the average and pays teachers badly — he put the average U.S. teacher salary at $30,000 a year — and that individualised digital curricula would be more effective, with robots held back until their interaction quality improves. 10

On October 10, 2025 — the day the host described as the largest liquidation event in crypto's history — Zhao's account is that a competitor pushed the story that Binance caused it, and he located the sequence of events in a tariff announcement and a market-wide decline rather than in his exchange. He also said a small technical issue with a stablecoin on Binance occurred and was not the cause. 10 The transcript used here is the publisher's automatically generated caption track, which garbles several proper nouns, so nothing beyond the passages quoted above is attributed to the recording.

Two other threads belong to the same week. On September 25, after Bitget disclosed that attackers had drained $351.6 million from its hot and warm wallets, Zhao wrote, "Tough day for Bitget. I expect and know @Binance, the @BNBCHAIN ecosystem, and the community will do everything we can to help. Stay SAFU!" 11 Bitget's chief executive said the attackers compromised a backend system in its wallet infrastructure, spoofed transaction data and triggered the authorization process, and that private keys were not compromised; the exchange said its User Protection Fund holds more than $464 million and covers the loss, kept deposits and trading open and suspended withdrawals. 38 That is the industry's real security story of the week: not a stolen key but a forged instruction that the approval system accepted.

On September 24 Zhao posted a photograph with the president of Sierra Leone, writing "Crypto adoption continues." 39 On September 23, answering a question about who runs his account, he said that for the past couple of years there has been no team and no automation and that all posts and replies are typed by hand. 40 That claim is worth noting only because it is unusual; a founder-typed timeline is a different kind of source from a corporate account, and it changes how much weight a reader should put on a one-line reply.

Vitalik Buterin, co-founder of Ethereum, published the window's most consequential text on September 27. "The cryptographic world computer," he wrote, is "my attempt to express in somewhat concise terms the true meaning of basically everything planned to happen to Ethereum starting from the fork after Hegota. It's really not just a blockchain anymore. It's a hybrid architecture that combines together blockchains and modern cryptography, to enable much more powerful properties. FOCIL, EIP-8288, Lean consensus, state management, formal verification, advanced mempool improvements (including privacy), and the longer-term specter of obfuscation all mentioned." 12

The essay's own page could not be retrieved in this environment — the site returned no readable body in either fetch mode — so what follows is attributed to the outlets that read it and to Buterin's own posts. 4142 As reported, the essay argues that post-Lean Ethereum is still called a blockchain largely for historical reasons, and that it merges the original ideas of the 2009 design with cryptographic tools that were unavailable or immature then. The fork named Hegota, planned for next year and combining an execution-layer upgrade called Bogotá with a consensus-layer one called Heze, is described as likely to be Ethereum's last "normal" fork; after it, recursive STARKs, formal verification and quantum-safe signature schemes take a larger role. Buterin projects slots of 4 to 8 seconds and finality of 8 to 32 seconds by 2030. He adds that Ethereum will not match the latency of centralised servers, and that decentralisation can be an advantage rather than a trade-off in specific places: more parallel data storage, more computation in parallel including inside the mempool, and better privacy because metadata about where data and requests originate is harder to observe. 41

Four adjacent posts from the same window show what he is treating as evidence for that claim. On September 26 he called PeerDAS — data availability sampling, which lets the network agree that block data exists without every node downloading all of it — an "underappreciated triumph", because it "has now been running for nearly a year with basically no problems": "the first large-scale instance of a blockchain that achieves consensus on the availability of data without needing any single node to download the entire block. Decentralized consensus without replication." 43 On September 27 he quoted a researcher's announcement of a formally verified decoupled consensus protocol that would speed up finality by four to eight times, with the comment: "Lean consensus will be amazing." The researcher's own summary argued that AI-assisted formal verification is now part of protocol design, because a design can be turned into a model, broken, and redesigned in a loop. 1344 On September 25 he flagged a new Solidity compiler. 45 On September 26, in a lighter register that is still the same argument, he listed what he would run on a local model — an Ethereum client, over 100 gigabytes of downloaded Wikipedia, Gutenberg and scientific articles — and called removing the need for centralised pinning services in IPFS a priority. 46

Two readers of the essay showed where it lands commercially. Stani Kulechov saw lending: "There are countless of use-cases where Ethereum verifiability would be useful beyond smart contact execution environment for finance to expand what we can actually do in DeFi while minimising trust." 47 Brian Armstrong quoted a longer thread arguing that if the vision is executed, what counts as "onchain" changes, because an application can push complexity offchain and inherit Ethereum's guarantees as long as the protocol accepts only outcomes that satisfy its rules — with the app-layer consequence that "a lot more of the smart stuff behind a financial product can happen outside the smart contract". Armstrong's own contribution was one word: "Interesting." 4849

The two checks are concrete. One is the schedule: Hegota is a multi-client fork, and a last "normal" fork announced a year ahead is the most falsifiable part of the plan. The other is where verification lands. Buterin has spent two years arguing that proofs about software are becoming possible; the test is whether formally verified components are running in a shipped protocol rather than in a model of one, and the consensus proposal he amplified on September 27 is the first item in the window that claims to be exactly that.

Arthur Hayes, chief executive of Flop Labs and chief investment officer of Maelstrom, published an essay on September 21 that argues the AI industry's caution is an economic event. "Safety First" opens by reading the AI labs' worry about humanity as a problem of unit economics: "cynical me looked at the calendar and thought, it's almost the close of the third quarter and Anthropic hasn't IPO'd yet." His claim is that the three leading U.S. labs' announcement that safety guides them to slow AGI development is not about human wellbeing but about a market that will not pay their prices, and that it matters far beyond AI because those labs' compute demand backs "over $1 trillion of investment grade and hundreds of billions of lower credit quality debt and loans." 14

The mechanism he describes runs like this. If the labs slow training spend, the debt raised against future chip and data-center purchases does not disappear; its price falls. The buyers of the safest tranches, on his account, are U.S. life insurers, some of them captives of the private-equity firms that also manage the debt — a captive insurer being one owned by its own investment manager, which then books affiliated reinsurance against its assets at a valuation its regulator keeps sealed. Hayes credits Nick Nameth of the Mispriced Assets newsletter with the mechanism and repeats Nameth's estimate of $1.54 trillion of affiliated reinsurance assets, while noting nobody knows how much of it is real. Policyholders are the ultimate holders of the risk, and the guarantee funds that would cover them are funded after an event rather than before. 14

His conclusion is that the government has two doors and both lead to the same room: either it becomes the buyer of compute of last resort in the name of national security, or it prints to keep the insurers solvent. "Either way, we as Bitcoin holders and crypto investors win." Along the way he assembled the money-supply argument he has been running for months: the Federal Reserve voted unanimously to raise its policy rate by 0.25%, its purchases of Treasury bills stopped on August 14, and commercial banks have since created over $100 billion of money by growing their total assets, while the rate rise hands them an extra $7.5 billion a year in interest on reserves parked at the Fed. His summary: "the conventional view that the Fed hiking rates and ceasing balance sheet expansion needs to include the effects on the commercial banking system as well. And when you combine the two together, the net effect is stimulative." He also noted the Fed's policy rate had not risen since July 2023, a detail that makes the hike a turning point rather than a repeat. 14 He announced the essay on X the next day with the shorthand "Trump has a choice, print or print." 50

His shorter posts were the same argument at market speed. On September 24 he watched the MOVE index, a gauge of expected Treasury-market volatility, rise past 130 and wrote, "The MOVE Index is surging in the right direction to force some sort of policy response. >130 and we are green lit for a bailout of some sort." 51 On September 23 he wrote that France's spread over German bunds was back at levels last seen in the 2011 euro crisis, with "Livre_Euro printer is warming up". 52

Flop Labs ran the week on agents. On September 23 Hayes reported that "over 4 million agents participated in our Technocore sonnet competition" and that the next contest would raise the prize tenfold. 53 By September 26 the trading competition had different numbers: "452k agentic traders, 8% have a positive pnl." 15 An 8% profitable share is not a failure for a competition that pays the top three, but it is a data point worth holding next to Brian Armstrong's argument that agents are the coming customer: most of these agents are losing, and the ones that are not are the ones being paid to compete. On September 27 Hayes amplified a teaser from an interviewer saying that Michael Saylor and Strategy "had their time in the sun, but the model is becoming less relevant as investors now have ETFs and other ways to get direct $BTC exposure", adding no words of his own; the interview itself had not been published when the window closed. 54

Two checks follow. The macro claim is arithmetic before it is opinion: Federal Reserve plus commercial bank total assets either keep growing or do not, and that series is public. The FLOP claim is about a ratio: testnet FLOP converts to mainnet FLOP at a rate the team has not published, and the next competition promises a million FLOP to be delivered once mainnet launches, which turns the conversion ratio into a price.

Andre Cronje, who founded Flying Tulip and created Yearn and Keep3r, published the cleanest numbers in the window for the question every token holder is asking about value accrual. "7 months since seed raise. 5 months since margin lending product launch. $100k protocol revenue. $1.44m put revenue. $1.34m in FT buyback and burns," he wrote on September 25. "FT has no vested tokens, no minting, no overhang, no token emissions. Just clean revenue." 16 The protocol's weekly update the same day reported fees past $100,000, premiums above 15% in the perpetual put marketplace, a new high in Lend total value locked, and ssftUSD, an auto-compounding wrapper that converts rewards into more ftUSD instead of paying them out. 55 The following day Cronje quoted his own post with two words: "Someone said buyback." 56

The ratio inside that scoreboard is the thing to look at. $1.44 million of put revenue against $1.34 million of buybacks and burns is roughly ninety cents spent per dollar earned, before the operating costs of a protocol that has been live for five months. That is a defensible growth posture and an unsustainable steady state, and the two are indistinguishable over one week. The compensation is that the revenue is real trading revenue rather than an emission, which is exactly the distinction the SEC staff's buyback answer drew when it separated buybacks from yield promises.

His second disclosure was a design choice that behaves like leverage even where it is not called that. Flying Tulip's lending market, he wrote, "uses equity based account valuation, not LTV," which lets a borrower reach effective loan-to-value ratios far above a position-level rule: his table lists USDC at 98.5% effective LTV, USDT at 98.5%, ETH at 83% and BTC at 83%, with margin-maintenance parameters of 150 basis points for the stablecoins and 1,700 for the two volatile assets. Liquidations, he added, are "not haircut based, but RFQ based, $ for $ repayments" — meaning the protocol asks market makers for a price on the collateral rather than applying a fixed discount. 17 A 98.5% effective LTV on a stablecoin leaves almost no room between the loan and the collateral, and a quote-based liquidation is only as good as the quotes available in the moment they are needed.

The window produced one podcast or interview with a tracked leader that can be sourced end to end: When Shift Happens with Changpeng Zhao, published September 25, whose publisher also posted the chapter list. 1036 Its transcript is an automatically generated caption track that misspells several proper nouns, so quotations here are limited to passages the caption renders unambiguously; the Hyperliquid, HYPE-listing and October 10 chapters are described only at the level of the publisher's chapter list. Arthur Hayes's interview with Wendy O was announced on September 27 with a single quoted claim, and the recording had not been published when the window closed, so nothing beyond that teaser is attributed to it. 54

Two first-party long-form texts landed inside the seven days: Hayes's "Safety First," read in full, and Buterin's "The cryptographic world computer." The second is the one that got away: the essay's page returned no readable body in either fetch mode, so its content is reported through the outlets that read it and quoted only where Buterin's own announcement post carries the words, and the essay's roadmap details are attributed to that reporting rather than to the text. 4142 No conference keynote from a tracked leader is verifiable inside this window. The week's primary documents are the SEC staff's September 25 FAQ, which no tracked leader linked without a commercial interest in the answer, and the Bitget breach disclosure, which no tracked leader other than Zhao addressed.