Bitcoin is testing the floor of its five-week range, hovering near 62,900 dollars after an intraday low of 62,538. A break below 62,500 could open the door to 58,500. Meanwhile, OpenAI dropped a bombshell for enterprises: a new Ultrafast mode running GPT-5.6 Sol at 14 times standard speed, powered by Cerebras hardware. Norm Law, the AI-native law firm, just opened a Washington DC office and is poaching partners from Big Law. A xAI co-founder pulled in 1.1 billion dollars for a new open AI stack called River. And Coldcard is dealing with the fallout from a 130 million dollar hack that has the entire self-custody world rethinking its assumptions. Let's get into it.
OpenAI and Cerebras just introduced Ultrafast, a new tier for GPT-5.6 Sol that spits out up to 750 tokens per second. That's roughly 14 times faster than the standard mode, and OpenAI is pitching it as a new speed class for frontier models. The benchmark they're showing off is telling: a 2,500-question evaluation that took 78 hours on a competitor's model finished in 11 hours on Sol Ultrafast, with similar accuracy.
What makes this work is hardware. Cerebras runs its wafer-scale engine with 44 gigabytes of on-chip SRAM, which eliminates the memory-movement bottleneck that plagues traditional GPU inference. It's a genuinely different architecture, and the speed gains show why specialized silicon matters when you're serving frontier models at scale.
The use cases OpenAI is highlighting are all real-time: incident response where you're parsing logs during an outage, financial research that reacts to market signals as they move, voice systems that handle complex customer issues without awkward pauses, and commerce flows that answer product questions mid-checkout. OpenAI is also using it internally for log analysis and faster research iteration.
Here's the catch. Ultrafast is invite-only preview. There's a waitlist. No general availability date. And critically, no published price. Standard Sol pricing is 5 dollars per million input tokens and 30 dollars per million output tokens, but the Ultrafast premium hasn't been disclosed. Cerebras capacity is the bottleneck, and OpenAI is being selective about which workloads get in.
One related move worth flagging: IBM just announced a deep partnership with OpenAI to embed GPT-5.6, Codex, and ChatGPT Work into IBM Consulting Advantage. They're launching a dedicated OpenAI Practice targeting regulated industries — finance, government, telecom, retail. If you're wondering how frontier AI actually lands inside a bank or a federal agency, this is the channel. The speed story and the enterprise-distribution story are converging fast.
The AI-native law firm is no longer a thought experiment. Norm Ai, the regulatory compliance platform, hit unicorn status this year on a 120 million dollar Series C at a 1.2 billion valuation. And its sibling entity, Norm Law, is now a real operating law firm competing directly with Big Law for institutional clients.
The pedigree here is what makes it credible. Norm Law is led by Mike Schmidtberger, former chair of Sidley Austin. They've hired partners from Kirkland and Ellis, Ropes and Gray, Proskauer, and Paul Hastings. As of this month they have 25 lawyers across eight practice areas — investment funds, real estate, commercial contracts, M&A, tech transactions, public filings, finance, and tax. And this week they opened a Washington DC office with two new partners focused on private funds and tech transactions. The client base already includes institutions managing more than 35 trillion dollars in assets.
The pricing model is the interesting part. Norm Law charges based on outcomes, not billable hours. The pitch is straightforward: Big Law's cost structure is misaligned with the value clients actually perceive, and AI-enabled workflows let a smaller team deliver high-value work at lower cost.
Elsewhere in the space, Scissero partnered with Mayer Brown on structured product issuance, targeting a market that generates over 100 million dollars in annual legal spend. GC AI acquired Jetify to scale its Contract Intelligence product, which runs on Anthropic's Claude Opus 4.6 and competes with Microsoft Copilot for in-house legal teams. And Neota Logic is pitching itself as the governance layer — every LLM call logged, checked against firm rules, routed for human sign-off, with full audit trails.
The pattern is clear. AI in law isn't just about faster document review anymore. It's about restructuring who does the work, how it's priced, and where the accountability lives. Big Law partners are voting with their feet, and the money is following.
Now the hard story. A firmware bug in Coldcard hardware wallets — reportedly five years old — let attackers drain roughly 130 million dollars in Bitcoin from thousands of addresses. The flaw bypassed the true random-number generator, making seed values predictable. Once you can predict a seed, you own the wallet. The bug was reportedly discovered by AI, which is its own uncomfortable footnote.
Here's what makes this incident so instructive. Coinkite's firmware is open source. It was public for five years. An AI-assisted security review did not catch the seed-generation flaw. So the comforting narrative that open source plus community audit equals safety just took a serious hit. Verification isn't automatic. It has to be actively performed, across repository boundaries, with reproducible builds and independent entropy checks.
The fix is uncomfortable. Coinkite issued patched firmware, but firmware can't repair an already-generated bad seed. Affected users have to migrate to entirely new seeds with fresh entropy. A strong BIP-39 passphrase helps but isn't a substitute.
The broader lesson: single-signer custody at scale is fragile. For meaningful holdings, the answer is multi-signature setups — say, 2-of-3 with keys from independent device families, so one compromised signer doesn't compromise the funds. Different vendors, different entropy sources, geographically distributed backups, and rehearsed recovery procedures.
Meanwhile the industry keeps moving. Ledger dropped a limited-edition Flex Hidden collab this week — 250 units, same underlying security stack. Breez launched Glow, an open-source web app that lets you send USDT and USDC directly from a Bitcoin balance using the Spark protocol and Lightning rails, with Passkey login and no seed phrase. MetaMask introduced Agent Wallet, giving AI agents on-chain trading capability with spending caps, protocol allowlists, transaction simulation, and up to 10,000 dollars per month in reimbursement for covered losses.
Self-custody isn't dead. But treating a hardware wallet as a magic box that solves the problem — that idea absolutely is.
On to the markets. Bitcoin heads into the weekend near 62,900 dollars, having touched 62,538 on August 14. That 62,500 to 62,560 zone has been the floor of a five-week trading range between roughly 62,000 and 66,000. Bitfinex analysts note the 65,000 to 65,500 region has rejected price six times. A weekend close below 62,500 could trigger a slide toward 58,500. Swan CEO Cory Klippsten thinks Bitcoin could bottom in October, roughly a year after the previous peak, and he's blunt about altcoins — his word was 'basically dead.'
Institutional flows tell a more interesting story. Second-quarter 13F filings just dropped and the picture is fragmented into distinct positional patterns. Morgan Stanley's IBIT holdings rose 23% to 16.5 million shares. Paul Tudor Jones added roughly 109,000 IBIT shares while cutting call equivalents by 85%, shifting from speculative options into direct spot exposure via the ETF. Abu Dhabi's sovereign funds — Mubadala and ADIC — held every single share of their combined 22.94 million IBIT position, absorbing a 118 million dollar mark-to-market hit without flinching. That's 764 million dollars of exposure that didn't blink.
On the corporate treasury side, the picture is uglier. One publicly disclosed Bitcoin treasury holding 1,145 BTC — worth about 67 million dollars — reported just 5,397 dollars in cash and a 7.4 million dollar working capital deficit. They need funding immediately or they're forced sellers. This is the pattern to watch for the rest of the year: leveraged treasuries with no operating cushion becoming liquidation risk during a drawdown.
And on the funding side of the market, capital is still moving. Velocity raised 38 million dollars Series A led by Dragonfly and FirstMark, with Coinbase and Ripple participating, building stablecoin treasury and settlement infrastructure for banks. River AI, from xAI co-founder Igor Babuschkin, pulled 1.1 billion dollars just two months out of stealth to build an open-weight AI stack that enterprises can own and run themselves. That last one matters because it points at the same theme cutting across every segment today — whether it's AI, law, or Bitcoin custody, the question keeps coming back to who actually controls the thing you depend on.
One question worth asking before you close the laptop today: how many of your critical dependencies — your custody setup, your AI stack, your legal exposure — are you actually verifying, and how many are you just trusting? The Coldcard bug sat in public code for five years. Verification is a discipline, not a default.