Editor’s Note

Three businesses spent this week finding out how much control they actually had, not how much they owned. OpenAI reminded Cursor that model access is a revocable privilege, not an asset, the moment its ownership changed. Sony Music and Warner Chappell reminded Anthropic that training on the world’s data comes with an invoice, and this week both of the industry’s remaining unaligned major publishers presented theirs at once. Tectonic learned that a lending market priced off a thinly traded token was never really under its control, and that gap cost an estimated $75 million.

None of these are new categories of risk: vendor lock-in, IP liability, and oracle manipulation have all been studied for years. What’s new is the density: three reminders inside three days, each landing on a company that had reason to think it held the leverage.

Meanwhile the actual leverage over crypto’s rally this week sat with Fed Chair Kevin Warsh, whose one hawkish line at a podium in Wyoming preceded a $3,700 drop in Bitcoin before lunch.

This issue covers the Cursor fallout, the widening AI copyright fight, a DeFi exploit that halted a chain, and the institutions building crypto infrastructure regardless.

Top Stories

1. OpenAI Cuts Cursor Off After SpaceX Buys It

The news. OpenAI notified SpaceX on August 28 that it will wind down the contract supplying models to Cursor, the AI coding platform SpaceX acquired for $60 billion in a deal that closed August 14. The cutoff lands November 12, the latest date OpenAI’s contract allows.

Why it matters. OpenAI’s own explanation is direct: “we cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk’s companies violating contracts,” citing Twitter’s past violations and Musk’s sworn admission that xAI distilled OpenAI’s outputs. It’s a company-specific dispute, and also proof a change-of-control clause can cut off access the moment your cap table changes, for reasons that have nothing to do with your product.

The take. OpenAI’s own logic holds together on its own terms, but the precedent is the real story. OpenAI has not alleged that Cursor itself breached anything; Cursor got acquired by a company OpenAI doesn’t trust, and lost its model supplier as a result. Every startup built on a foundation-model API now has a concrete example of how fast that access can vanish. More in this week’s Deep Dive.

2. Three Major Music Publishers Are Now Suing Anthropic

The news. Sony Music Publishing and Warner Chappell Music filed suit against Anthropic on August 28 in California, alleging a “brazen campaign” of torrenting, scraping, and downloading copyrighted lyrics, then stripping song titles, writer credits, and copyright notices before training Claude on the material. Cited works include “Eye of the Tiger,” Marvin Gaye’s “Ain’t No Mountain High Enough,” and Taylor Swift’s “Paper Rings.” The complaint names CEO Dario Amodei and co-founder Benjamin Mann personally as defendants, alongside the company, and seeks statutory damages up to $150,000 per work infringed plus $25,000 per stripped copyright notice.

Why it matters. Universal Music Publishing sued Anthropic on similar grounds in 2023. Sony and Warner joining puts all three major music publishers in active litigation against the same company at once, for the first time. It also lands about six weeks after a judge finalized Anthropic’s $1.5 billion settlement in the unrelated Bartz author case, where Judge Alsup ruled that training on copyrighted books is fair use but downloading them from piracy sites is not, then Anthropic settled rather than let the piracy question go to trial. Sony and Warner are making the same accusation again, aimed at lyrics instead of books. An Anthropic spokesperson told TechCrunch the company disagrees and intends to “defend ourselves robustly in court.”

The take. Bartz drew a distinction that will loom over this case: training is fair use, piracy is not. But that’s one district judge’s ruling, not settled industry law; Anthropic’s decision to settle kept the piracy question from ever reaching an appeals court. The wrinkle that might matter more is the allegation Anthropic stripped copyright-management information before ingestion, a separate DMCA question that turns on facts not yet public. Worth watching, not yet worth calling more dangerous than the piracy claim, which is what Anthropic already paid $1.5 billion to settle once.

3. DeepSeek’s $7.4 Billion War Chest Sets Up a Public Listing

The news. DeepSeek is closing in on $7.4 billion in new funding at a $74 billion pre-money valuation, the Wall Street Journal and South China Morning Post reported, up from just over $50 billion in June. Bloomberg had reported the round in earlier talks as up to $8 billion; it appears to have settled near $7.4 billion as it closed. Backers Monolith, Shixiang Capital, and CATL are participating, with the round expected to finalize by end of August ahead of a targeted 2027 Shanghai IPO.

Why it matters. The equity structure reportedly changed with the price: investors who previously held stakes indirectly through a limited partnership run by founder Liang Wenfeng can now invest directly, a shift read as pre-IPO housekeeping. BigGo Finance, citing people familiar with DeepSeek’s financials, put annualized recurring revenue around $500 million against the $74 billion valuation, a roughly 148x multiple that only works if investors expect DeepSeek to keep taking share from OpenAI and Anthropic, not just survive next to them.

The take. A Shanghai listing would give Chinese capital a direct equity path into frontier AI inside China’s own regulatory perimeter, something no U.S. lab currently offers. What that 148x multiple is really pricing is compute access holding up under export controls. Watch the IPO filing timeline as the tell; a slip would say more than any funding headline.

4. A Twenty-Minute Price Pump Halted an Entire Blockchain

The news. Cronos, the blockchain tied to Crypto.com, halted block production on August 30 after an attacker pumped TONIC, the thinly traded governance token of lending protocol Tectonic, roughly 100x higher in about twenty minutes, then borrowed against the inflated collateral. Onchain researcher Weilin Li estimates the exploit drained roughly $75 million from Tectonic’s $121.7 million in total value locked; only about $6 million reached Ethereum before validators froze the chain.

Tectonic total value locked, August 26 vs. August 30, 2026

Why it matters. This is a Mango Markets-style oracle manipulation attack, well documented since 2022, and it’s the second such attack this week: Moonwell lost $8.7 million to the identical mechanic on Base three days earlier. The playbook isn’t new. What’s notable is a leading lending protocol on an exchange-linked chain still had no defense against it in 2026.

The take. Halting the chain is blunt: it stopped an estimated $69 million from leaving Cronos, but it’s also a centralized intervention on a network marketed as decentralized. That tradeoff keeps recurring because most chains still don’t have a better answer than pause and hope. Crypto.com’s app and exchange were unaffected, which will matter more to Cronos’s reputation than the exploit itself.

5. Fed Hawkishness Snaps Crypto’s Rally

The news. Bitcoin touched a three-month high near $81,455 on August 28, then fell sharply to about $77,678 by the next day, in the hours after Fed Chair Kevin Warsh’s hawkish Jackson Hole speech that Friday morning. Warsh said the Fed has “work to do” on inflation, and September rate-hike odds rose from 35% to 42% on CME FedWatch in the hours that followed. Bitcoin has since stabilized near $78,700-$79,000.

Bitcoin price, August 26 to 31, 2026

Why it matters. Last issue we covered large wallets accumulating through the rally rather than chasing it, consistent with large-holder or institutional accumulation rather than retail euphoria (some of those wallets are exchanges or custodians, not single buyers). This week is the other half of that picture: accumulation doesn’t mean immunity. Bitcoin remains fully exposed to conventional monetary policy, no matter who’s holding it.

The take. A nearly $4,000 move in under 24 hours, in the window right after one hawkish speech about inflation, reinforces how sensitive Bitcoin remains to monetary-policy expectations, whatever the “uncorrelated digital gold” pitch claims. Watch whether wallets holding 10,000+ BTC keep accumulating through this pullback or start trimming; that will be a cleaner read on conviction than the price alone.

6. Sberbank Prepares to Take Bitcoin as Loan Collateral

The news. Sberbank, Russia’s largest bank, plans to accept Bitcoin as loan collateral as the country’s new digital-asset law, signed by President Putin on August 4, takes effect September 1 under Bank of Russia oversight. Deputy chairman Anatoly Popov told Russian state news agency TASS that Ether and USDT follow once the central bank permits their public circulation. Sberbank issued what it called Russia’s first bitcoin-backed loan to miner Intelion Data in December.

Why it matters. Crypto payments remain banned domestically under the same law; this is collateral, not currency. That split is the whole story: Russia’s largest, majority state-owned bank is building institutional crypto infrastructure while keeping crypto out of the payment system entirely, a distinction most Western regulators have avoided stating so explicitly.

The take. Treat crypto as collateral first and a payment rail never, and let institutional custody build before any consumer use case follows. The catch is USDT: Tether can freeze tokens tied to sanctioned entities, and Sberbank has been under full U.S. and EU sanctions since 2022, making stablecoin collateral a different risk category than Bitcoin or Ether, whatever the Bank of Russia’s draft list says. Watch which asset the bank actually books first.

Quick Hits

Google’s A2A protocol formally joined the Agentic AI Foundation, the Linux Foundation body already hosting Anthropic’s Model Context Protocol, which has grown from 49 founding members to more than 250 in under a year. MCP handles agent-to-tool; A2A handles agent-to-agent.

IBM and the University of Chicago demonstrated a verified quantum computation beyond the reach of leading classical simulation methods, using 70 error-corrected logical qubits to finish a task in roughly 15 minutes.

XRP treasury vehicle Evernorth cleared its SEC registration on August 27. Shareholders vote September 30 on a Nasdaq listing under ticker XRPN, backed by roughly 473 million XRP.

OpenAI retired the standalone DALL-E GPT inside ChatGPT on August 30. Image generation isn’t going away, it moves to ChatGPT Images, the newer tool that was already most users’ default.

Circle and USDC become Chelsea FC’s front-of-shirt sponsor starting the 2026/27 season, crypto’s most prominent push into mainstream European football.

Your AI Stack Can Be Revoked

Read past the Musk feud in this week’s Cursor story and there’s a more uncomfortable fact underneath: Cursor lost its model access not for anything OpenAI has alleged it did, but because its ownership changed, and a standard contract clause turned out to carry outsized consequences. That should worry a wider set of founders than the ones building coding tools.

The clause was always there

OpenAI’s own explanation is candid about the mechanism: “our custom agreement with Cursor gives us a limited time window to cancel it after a change of control.” Change-of-control termination rights are standard in enterprise software generally, so a vendor isn’t forced to keep serving a customer that becomes, through acquisition, a different and potentially unacceptable counterparty. Nothing about that is unusual.

What’s different is degree, not kind. Migrating off a CRM after such a clause fires is a real cost, but the data usually survives the move. Migrating off a foundation model you’ve fine-tuned and built a context window’s worth of product behavior on top of is a harder swap, and for some products, not yet a possible one on the day the notice period ends.

Nobody outside the deal knows how many of these clauses exist

These are private contracts. We don’t know how many AI-native startups carry similar language, or how aggressively any lab would enforce it against a less politically charged acquirer than SpaceX. Cursor’s situation is legible only because Musk’s history with OpenAI is well documented and OpenAI chose to publish its reasoning. A quieter acquisition could trigger the same clause with zero public explanation.

The obvious objection

Labs need this power. If a model ends up serving an acquirer with a record of misuse, or one that raises national-security questions, the lab is the only party positioned to cut off access before harm occurs, and that argument gets stronger as frontier capability grows. That’s reasonable, and it’s also true that it protects the lab’s interests first. Nothing requires those two to be the same interest.

What this means for founders

Treat model access as a covenant with a counterparty, not a fixed input. Audit change-of-control language in your own contracts now, before an acquirer or new investor makes it relevant. Where the product allows it, maintain a credible fallback to a second provider, even a degraded one, so a vendor’s decision about someone else’s cap table can’t zero out your product overnight.

Disclosure: Building Blocks holds no position in OpenAI, Anthropic, Cursor, or SpaceX, and received no compensation for this piece.

Sources

Top Stories: Our decision on Cursor following its acquisition by SpaceX (OpenAI) · OpenAI to end model access to Cursor after acquisition by Elon Musk’s SpaceX (CNBC) · Sony Music, Warner sue Anthropic, alleging a “brazen campaign” of intellectual property theft, incl. Anthropic’s statement (TechCrunch) · Sony Music Publishing and Warner Chappell sue Anthropic in multi-billion dollar lawsuit (Music Business Worldwide) · Anthropic’s landmark $1.5B copyright settlement is approved (TechCrunch) · DeepSeek Nears $7.4 Billion Funding Round at $74 Billion Valuation, citing WSJ and SCMP (China Money Network) · DeepSeek Resumes $8 Billion Round With Monolith in the Running (Bloomberg) · DeepSeek Reportedly Launches New Funding Round, revenue detail (BigGo Finance) · Crypto.com-linked Cronos network halts after Tectonic exploit estimated at $75 million (The Block) · Warsh at Jackson Hole: ‘We have work to do’ on inflation (CoinDesk) · Sberbank plans to add ether and USDT as collateral for crypto-backed loans (CoinDesk) · Sberbank issues Russia’s first crypto-backed loan to bitcoin miner Intelion Data (CoinDesk)

Quick Hits: Google’s A2A Protocol Joins AAIF (Forkast) · IBM and The University of Chicago Demonstrate Quantum Advantage (IBM Newsroom) · An XRP treasury company backed by Ripple is a shareholder vote away from Nasdaq (CoinDesk) · Evernorth sets Sept. 30 vote for XRP treasury merger (crypto.news) · You have until August 30 to save your ChatGPT DALL-E images (Tom’s Guide) · Circle strikes Chelsea front-of-shirt sponsorship deal featuring USDC (The Block)

Closing

The most valuable clause in your stack this week wasn’t a security patch or a price feed. It was the one buried in a contract nobody expected to matter until an acquisition made it the whole story. Read your own before someone else’s decision reads it for you.

Learn. Connect. Build. Repeat.

Building Blocks

Editor’s Note

Three things happened this week that look unrelated until you put them side by side. The White House convened crypto executives and its own regulators to route policy around a stalled Congress. Nvidia structured a $6 billion deal that looks nothing like an acquisition, on paper. A nonprofit graded five frontier AI labs on stopping a rogue model, and found none scored above “partial implementation” on anything.

The pattern: everyone involved, from Washington to Menlo Park, is building around scrutiny rather than through it. Regulators are doing it deliberately, carving out safe harbors because the old rules never fit crypto’s economics. Dealmakers are doing it opportunistically, structuring transactions to stay one signature short of a merger filing. AI labs are doing it by omission, disclosing just enough to avoid a failing grade.

None of this is illegal. Most of it might even be reasonable on its own terms. But stack it together and the picture is an industry getting good at looking compliant while keeping the mechanics of power (who owns what, who controls what) mostly out of view.

This issue covers all three threads, plus the whales quietly buying the dip everyone else is calling a retail rally.

Top Stories

1. Congress Isn’t Writing the Crypto Rules Anymore

The news. The White House held a closed-door crypto summit on August 19 with Coinbase, Ripple, Gemini, Robinhood, Kraken, and CFTC and SEC leadership. It landed a day after the SEC’s Regulation Crypto Assets proposal ($5 million and $75 million exemptions, plus a safe harbor ending securities status once “essential managerial efforts” are complete) and two days after Treasury’s GENIUS Act stablecoin rule, both now open for comment.

Why it matters. The CLARITY Act, the market-structure bill the industry actually lobbied for, faces a cloture vote on September 15 needing 60 votes. Polymarket odds of passage have fallen from 82% in February to under 20%. Rather than wait on Congress, the administration is using the SEC and Treasury to write the rules directly.

The take. The tell is the CFTC’s new Innovation Advisory Committee, seated the day after the summit with a roster weighted toward the CEOs of Polymarket, Kalshi, Coinbase, Robinhood, and DraftKings, and no consumer advocates. That is a drafting body, not an oversight one, and the SEC’s proposing release reads like the opening move in the same plan, not an isolated rulemaking.

2. The Rally Isn’t Retail’s

The news. Bitcoin posted its best five-day run since March 2024, crossing $79,000 before easing to roughly $77,200. XRP gained close to 40% on the week and Zcash jumped over 30%. Bitcoin ETFs pulled in $1.9 billion over the week, including $307 million on August 21 alone.

Why it matters. The number worth isolating: wallets holding at least 10,000 BTC climbed to 90, a six-month high, even as smaller retail wallets reduced their holdings into the rally. That is accumulation, not euphoria, and it changes how durable the move likely is.

The take. Retail-driven rallies tend to reverse hard once the short squeeze clears. Accumulation into strength is stickier: large holders expect the move to continue rather than fade. Analysts are right that Bitcoin needs to hold above $70,000 to prove this outlasts the squeeze. The wallet data says the people with the most at stake already believe it will.

3. Nvidia’s Poolside Deal Is an Acquisition in a Costume

The news. Nvidia agreed to pay AI coding startup Poolside $6 billion to license its Model Factory training system and make offers to 109 of its staff, plus a separate $1 billion investment at a $12 billion pre-money valuation. Poolside’s three co-founders stay in place, and the company keeps operating independently, at least on paper.

Why it matters. This is the second deal of its kind this year, after Nvidia’s similar arrangement with Groq. The structure lets Nvidia absorb a startup’s core technology and talent without filing the merger notification that would trigger antitrust review, since no controlling stake technically changes hands.

The take. Call it what it functions as: a reverse acquihire, engineered to stay one signature short of a transaction regulators would examine. Newcomer’s reporting lays out the mechanics, and they mirror the Groq deal closely enough to look like a template. More in this week’s Op-Ed.

4. No AI Lab Can Say How It Would Stop a Rogue Model

The news. GuideLight AI Standards’ scorecard on five frontier labs’ containment practices drew wide coverage on August 22, days after publication. Anthropic and OpenAI tied for the top score at C+, Google scored a D+, xAI a D-, and Meta an F. No company scored above “partial implementation” on any of the six practices measured, including containing a model misbehaving at scale.

Why it matters. These grades measure public disclosure, not internal capability. A lab could have a real containment plan and still score an F for not publishing it. But that gap is itself the finding: the industry building uncontrolled AI agents decided the public doesn’t need to see any containment plan.

The take. GuideLight’s methodology is transparent about what it can and cannot measure, more than most safety scorecards manage. OpenAI’s 3-out-of-5 on containment, the group’s highest, credits it for pausing workloads after past incidents. That’s a real signal, and still a passing grade on a test where the industry median rounds to failing.

5. ChatGPT Ads Land in Europe, Carefully

The news. OpenAI began serving ads inside ChatGPT across 31 European markets on August 24, six months after the U.S. pilot launched in February. Free and Go users see ads; paid tiers stay ad-free. European ads launch without personalization, targeted only on the current conversation, approximate location, device type, and language.

Why it matters. This is OpenAI’s first ad rollout into GDPR territory, and no-personalization is a concession to European privacy law, not a product choice. It also lands alongside confirmation that OpenAI’s enterprise revenue has overtaken consumer revenue, making ads look more like a hedge against churn than a growth engine.

The take. The interesting number isn’t 31 markets, it’s the six-month gap between the U.S. pilot and the European rollout. That is roughly how long it took OpenAI’s legal team to build ad targeting that survives GDPR scrutiny. Expect that same compliance-first playbook wherever OpenAI expands next, especially into markets with U.S.-style privacy rules pending.

Quick Hits

Google ties itself to Marvell’s TPU roadmap. Marvell granted Google warrants for up to $12.2 billion in shares, vesting as Google buys TPU chips, a deal that could make Google a top-five Marvell shareholder if $120 billion in orders lands by 2033.

Grayscale is renaming its Zcash Trust “The Zcash ETF.” Its fifth SEC amendment, filed August 21, sets up a NYSE Arca listing around August 25, with a Digital Currency Group subsidiary in talks to contribute nearly 200,000 ZEC.

A Canadian firm made post-quantum cryptography boring, in a good way. Crypto4A’s QASM module became the first hardware security module to earn FIPS 140-3 Level 3 validation supporting every NIST-standardized post-quantum algorithm, on August 20.

TRON’s Super Representatives vote August 25 on a compatibility upgrade. GreatVoyage-v4.8.2 adds a CLZ opcode and secp256r1 verification, both long-standard on Ethereum, and needs approval from 27 active Super Representatives to go live.

The Reverse Acquihire Is Regulatory Arbitrage, Not Innovation

Nvidia’s $6 billion deal with Poolside is being covered as a licensing agreement. It is not one, functionally. Poolside hands over its core training system, offers jobs to the engineers who built it, and takes a $1 billion check tied to a valuation Nvidia effectively set. Call it a licensing deal if the paperwork says so. Treat it as an acquisition, because that is what actually changes hands.

This is not a one-off. Nvidia used a nearly identical structure with Groq earlier this year: license the technology, hire the team, invest at a valuation nobody outside the deal can verify, leave the org chart technically intact.

Two deals with the same shape is a pattern, and a pattern this clean, from a company this large, deserves to be named for what it is: a way to acquire market power without filing the paperwork that would let anyone examine it.

Why the paperwork matters

Merger review exists because concentrated control over critical inputs, chips, foundation models, distribution, tends to produce worse outcomes downstream: higher prices, slower innovation, fewer independent labs able to compete. The Hart-Scott-Rodino process forces companies above a size threshold to disclose a deal before closing it, giving regulators a chance to object.

Nvidia’s structure with Poolside was built to sit just outside that trigger. No controlling stake changes hands, no board seats change hands, and on paper Poolside remains independent. In practice, Nvidia now controls the technology Poolside was built to develop and employs the people who built it. The independence is nominal.

Newcomer’s reporting on the deal’s mechanics describes exactly this: a licensing fee and a hiring spree engineered to look like neither an acquisition nor an acquihire, because either label would invite scrutiny the structure avoids.

The obvious objection

The obvious objection is that this is efficient dealmaking, not evasion. Poolside gets capital, and its investors get a payout years before an uncertain IPO. Nvidia gets talent and technology without the multi-year drag of a full acquisition. Nobody is coerced. The three co-founders chose this. If the outcome is good for both sides, why does the label matter?

It matters because merger review protects more than the parties to a deal. It protects everyone not in the room: rival labs facing a better-resourced competitor, customers losing an independent option, and future startups that will default to this structure because it worked.

What regulators should actually do

The FTC and Department of Justice have the tools to close this gap. Hart-Scott-Rodino’s thresholds are based on transaction value and party size, not deal structure. Nothing stops regulators from writing guidance that treats an exclusive license paired with a mass-hiring offer as the functional equivalent of an acquisition for reporting purposes. The tool exists. It has not been pointed at this structure yet.

Disclosure: Building Blocks holds no position in Nvidia, Poolside, or Groq, and received no compensation for this piece.

If the Poolside deal goes unchallenged, expect two or three more before year-end structured the same way, aimed at mid-sized labs sitting on technology too valuable to ignore and too small to survive a fair fight for talent. Watch OpenAI and Anthropic’s smaller competitors for the next one.

Sources

Top Stories: The White House Summit Isn’t a Photo Op (Yahoo Finance) · Trump pushes Congress to move on Clarity Act (CoinDesk) · Treasury Seeks Public Comment on GENIUS Act Proposed Rulemaking (Treasury.gov) · Polymarket CLARITY Act odds crashed from 82% to under 20% (Yahoo Finance) · SEC Proposes New Regulation Crypto Assets (SEC.gov) · XRP leads altcoin rally, Bitcoin’s biggest weekly gain in two years (The Block) · Poolside strikes $6 billion licensing deal with Nvidia (Newcomer) · Nvidia to pay $6 billion in licensing and hiring deal (The Information) · Frontier AI labs still won’t say how they’d contain a rogue model (TechCrunch) · Guidelight’s Control Assessment of Frontier AI Companies (GuideLight) · OpenAI to launch ChatGPT ads in 31 European markets (Dataconomy) · ChatGPT ads reach Europe Monday (Tech Times)

Quick Hits: Google secures $12.2 billion share purchase right in Marvell deal (Bloomberg) · Marvell jumps on $12.2 billion AI chip deal (CNBC) · Grayscale files fifth SEC amendment for Zcash ETF (CoinPaprika) · Grayscale files fifth SEC amendment for proposed Zcash ETF (Crypto Briefing) · Crypto4A sets global benchmark with FIPS 140-3 Level 3 validation (PR Newswire) · TRON prepares for TVM compatibility upgrade, voting set for August 25 (Crypto Briefing)

Closing

The SEC’s comment period runs 60 days. Most of the industry will spend it arguing about the dollar caps. The fight worth watching is over four words: “essential managerial efforts.” Whichever way that phrase gets defined by mid-October decides whether decentralization becomes a legal finish line or stays a marketing claim.

Learn. Connect. Build. Repeat.

Building Blocks

Welcome to Building Blocks

You are reading the first edition of Building Blocks Weekly, a newsletter for the people building, funding, and operating the next layer of frontier technology: founders, investors, operators, and developers working in blockchain, AI, and frontier tech.

Every Tuesday, one edition, built around a single principle: signal over noise. We read what the industry reads, then throw out whatever will not matter in six months and write up what is left.

Here is what to expect each week.

• Editor’s Note. The macro narrative connecting the week’s noise.

• Top Stories. The developments that will shape the next quarter, with our honest read on why each matters.

• Quick Hits. Short, high-signal updates worth knowing but not dwelling on.

• Deep Dive. Original analysis on one theme, contrarian where warranted.

• Monthly Feature. Once a month, a video interview with a builder, founder, or investor we trust.

We do not write hype or clickbait, and we do not pretend to certainty we do not have. Where we have an opinion, we say so. Where we are guessing, we tell you.

If an edition is useful to you, forward it to one person you respect. That is how this grows. Operators talking to operators.

Now, on to this week.

Editor’s Note

Two industries got a stress test this week, and neither one passed cleanly.

On August 14, Anthropic raised its own catastrophic-misalignment rating from “very low” to “low,” citing a wave of recent disclosures about how frontier models behave inside cybersecurity evaluations, not any new failed test.

Two weeks earlier, Bitcoin owners who did everything self-custody recommends had already started losing what is now $116 million to a firmware bug sitting in their wallets since 2021. Neither failure was exotic: a company and a piece of code quietly falling short of what users assumed was true.

The capital keeps arriving regardless. Stripe agreed this week to pay more than $7 billion for OpenRouter, a 5.4x markup on a valuation set three months ago. DeepSeek, having built its reputation on being the cheap option, just raised its API prices by as much as 1,100%.

Bitcoin ETFs pulled in $853 million in a single week. None of that money is waiting for the safety model to catch up with the growth model. That gap is this week’s real story.

This issue covers where the money is moving, where the guardrails are not, and one argument for why self-custody’s real vulnerability was never the exchange.

Top Stories

1. Stripe Just Bought Its Way Into the Agentic AI Economy

The news. Stripe agreed to acquire OpenRouter, the API gateway that routes developer traffic across more than 400 AI models, for more than $7 billion, a 5.4x markup on the $1.3 billion valuation OpenRouter set in May 2026.

Why it matters. OpenRouter is the most expensive in a string of acquisitions Stripe has made over eighteen months to build itself into agentic AI’s financial infrastructure layer. Whoever bills the agent economy sits underneath every transaction it generates.

The take. A 5.4x markup in three months is a strategic bet, not a valuation update. Stripe decided owning the routing layer, the point where a request gets matched to a model and a price, is worth more than fundamentals suggest. Model routing was infrastructure. Stripe just made it a moat.

2. Anthropic’s Profit Is Up. So Is Its Own Risk Estimate.

The news. Anthropic reported preliminary Q2 revenue above $11.5 billion, up 130% from Q1, and its first quarterly operating profit, roughly $559 million. Days later, on August 14, it raised its own catastrophic-misalignment rating from “very low” to “low,” citing uncertainty from recent cybersecurity-evaluation disclosures rather than a new failed test.

Why it matters. Anthropic moved its own risk dial in the wrong direction voluntarily, in the same window it announced beating its profitability roadmap by two years.

The take. Financial beats get restated when they prove too conservative. Safety ratings almost never do, since raising your own risk number in public cuts against the commercial incentive. Whether this is caution or expectation-setting ahead of Model 2, the stronger unreleased model the same report disclosed, is worth watching.

3. A Firmware Bug From 2021 Just Cost Bitcoin Owners $116 Million

The news. A vulnerability in Coldcard hardware wallet firmware, live since version 4.0.0 in March 2021, caused devices to generate seed phrases from a predictable substitute instead of the dedicated randomness chip. Since July 30, an attacker has drained roughly 1,816 BTC ($116 million) from more than 5,200 addresses, including 1,083 BTC ($70.2 million) in one 41-minute sweep.

Why it matters. Every wallet generated on an exposed Coldcard between March 2021 and the patch is compromised, whether or not the device or PIN was touched. Self-custody removed the exchange as a point of failure and replaced it with the firmware supply chain.

The take. Bitcoin developer James O’Beirne published an analysis on August 4 linking the pseudonymous account behind the weak code to Coinkite CTO Peter Gray, via matching signatures across commits. He also says he privately flagged the defect to Coinkite in May 2025 and was ignored. Coinkite has not responded. The Deep Dive below argues that, if accurate, should change how the industry talks about self-custody.

4. Stablecoins Are Getting Regulated Everywhere Except Where They Started

The news. The GENIUS Act’s July 18 deadline for finalized U.S. stablecoin rules passed without a coordinated package from the six federal agencies involved, though comment windows run into August. The EU, UK, Singapore, Hong Kong, UAE, and Japan now mandate full reserve backing, licensed issuers, and guaranteed redemption.

Why it matters. The country that hosts the two largest stablecoin issuers by market capitalization is currently the one without a finished rulebook, while every other major financial center has either finished theirs or is close.

The take. This is not America falling behind on innovation. Issuers already operate under de facto federal comfort and state licenses. It is a story about who writes the rules overseas issuers will eventually match, and that pen sits on six desks that have not agreed to pick it up together.

5. Wall Street Ran a Live-Fire Drill on Tokenized Markets

The news. Close to 40 firms, including JPMorgan, Goldman Sachs, Invesco, and Citadel Securities, spent July trading tokenized stocks and Treasuries over blockchain networks, while the DTCC monitored from war rooms in New York and New Jersey. Goldman separately agreed to acquire NEOS Investments, a $30 billion options-income ETF manager, for up to $2.25 billion.

Why it matters. This was not a hackathon. The DTCC, which clears most U.S. securities trades, watched whether tokenized settlement can match the risk controls of the system it already runs.

The take. The NEOS deal and the tokenization drill are the same bet from two directions: yield products retail investors already trust, and settlement rails that might make moving between products cheaper. Watch whether the DTCC publishes findings. A quiet no-comment would say more than a press release.

Quick Hits

• DeepSeek just abandoned the price war it started. V4-Pro launched August 13 at roughly 9 to 14 times V4 Flash’s price, and a broader V4 price hike of up to 1,100% took effect August 16.

• Binance is cutting off eleven platforms, including HTX, over sanctions. The move follows HTX’s inclusion in the EU’s Russia sanctions package and a UK designation of Huobi Global, not a unilateral Binance decision.

• A US bank regulator opened the door for World Liberty. The OCC granted preliminary conditional approval on August 14 for the Trump family-linked venture’s trust company to organize as a national bank.

• Cboe wants the first 3x leveraged Bitcoin and Ether ETFs. Volatility Shares filed August 10 for futures-based funds, which need individual SEC approval since they fall outside standard leveraged-listing rules.

• Russia’s central bank approved retail Bitcoin, Ethereum, and USDT trading. The framework caps purchases at 300,000 rubles a year per intermediary and takes effect September 1.

Your Keys, Someone Else’s Firmware

Bitcoin’s founding promise was simple: hold your own keys and no counterparty can lose your money for you. Mt. Gox, FTX, and a decade of exchange collapses turned that promise into an industry-wide slogan. “Not your keys, not your coins” is printed on t-shirts because it is true.

It is also incomplete, and the Coldcard hack made that plain. More than 5,200 Bitcoin owners did exactly what the slogan told them to do: bought a hardware wallet, generated a seed offline, never touched an exchange. They still lost $116 million, because the device that was supposed to remove counterparty risk was itself a counterparty.

Coinkite wrote the firmware. A code path shipped in March 2021 skipped the dedicated randomness chip, a defect confirmed by the patch itself, whatever its authorship. Nobody who lost funds made a mistake a t-shirt could have prevented.

This is not an argument against self-custody. Exchange custody has produced far larger losses over a longer period, and nothing about the Coldcard incident changes that base rate. The argument is narrower: self-custody trades a legible counterparty you can evaluate, an exchange with a balance sheet, a jurisdiction, a track record, for an illegible one.

Almost nobody checks a hardware wallet’s firmware commit history before buying it. Almost nobody could, even if they tried. James O’Beirne’s August 4 analysis alleges the weak code traces to Coinkite’s own CTO, Peter Gray, writing under a pseudonym, and says he warned Coinkite about it in May 2025. Coinkite has not responded.

If that holds up, the counterparty risk here was not a hacker. It was an engineering decision inside the company whose entire product pitch is that you do not have to trust anyone, one it was reportedly told about more than a year before it cost customers $116 million.

The obvious objection is that this is what open source is for. Coldcard’s firmware is publicly auditable, and the bug was eventually found and patched, which is more than can be said for a black-box custodian’s internal controls. That is fair, and it is also five years late for the 5,200 people whose seeds were already generated.

Auditable is not the same as audited. Open source shifts the possibility of scrutiny onto the public. It does not guarantee anyone with the skill and time actually performed that scrutiny before March 2021.

The second objection is sharper: exchange custody concentrates risk into a single honeypot, while a firmware bug like this one requires an attacker to know which addresses came from a Coldcard and target them individually. That is true today.

It stops being true the moment a weak random number generator becomes a known, scriptable exploit, which is roughly where this bug sits now. Concentration risk and supply chain risk are different shapes of the same problem: a few decisions, made by a few people, that most users cannot independently verify.

What should change is not the self-custody recommendation. It is the messaging around it. “Not your keys, not your coins” implies that holding your own keys ends the trust question. It is the start of a different one: whose firmware, whose supply chain, whose code review process.

Hardware wallet vendors should publish independent, dated security audits the way custodians publish proof-of-reserves. Buyers should ask for the audit before the marketing copy.

The Coldcard hack will not be the last firmware-level failure in hardware custody, because the incentives that produced it, small teams, thin audit budgets, and a customer base that trusts the hardware precisely because it does not understand it, have not changed.

If a comparable bug surfaces in a competing wallet before the end of 2026, treat this column as the baseline, not the outlier. That is a dated forecast, and readers are welcome to hold us to it either way.

Disclosure: Building Blocks holds no position in Coinkite, Coldcard, or any competing hardware wallet manufacturer.

Sources

Top Stories: Stripe to acquire OpenRouter (Bloomberg) · Anthropic Risk Report, August 2026 · Anthropic Q2 operating profit (Let’s Data Science) · Coldcard exploit breakdown (TRM Labs) · Coinkite CTO linked to Coldcard code (Cryptopolitan) · Global stablecoin regulation tracker (BVNK) · Wall Street tokenization drill (Bloomberg) · Goldman Sachs to acquire NEOS Investments

Quick Hits: DeepSeek V4-Pro pricing (Caixin Global) · Binance halts HTX transactions over sanctions (Crypto.news) · World Liberty wins OCC preliminary approval (Bloomberg) · Cboe files for 3x leveraged Bitcoin and Ether ETFs (The Block) · Russia approves retail crypto trading framework (The Moscow Times)

Closing

Nothing that went wrong this week was actually unforeseen. Anthropic named its own risk before anyone forced it to. O’Beirne says he flagged the Coldcard defect fifteen months before it drained a wallet.

The open question for next week is not whether more capital finds its way into crypto and AI. It clearly will. It is whether the guardrails get built before or after the next foreseeable failure lands, and whether Coinkite ever answers O’Beirne directly. That silence is already a data point.

Learn. Connect. Build. Repeat.

Building Blocks

Share Building Blocks