Bitcoin punched through $81,000 overnight before pulling back, its highest level since May. Miner stocks are ripping — Canaan, American Bitcoin and Cango jumped as much as 67% on the week. Meta dropped open weights for a new 30-billion-parameter model called Muse Glimmer that runs on a single consumer GPU. Salesforce and Anthropic launched Claudeforce, embedding the entire CRM inside Claude. And the US Treasury finally proposed licensing rules under the GENIUS Act, setting a January 2027 deadline for stablecoin issuers. Let's get into it.
Bitcoin touched $81,455 overnight, its highest print since May 15. The move came alongside a Nvidia earnings beat, renewed regulatory optimism after a White House meeting with crypto executives, and a brutal short squeeze that liquidated over $1.6 billion in crypto positions in 24 hours. Coinbase premium flipped positive for the first time since May — meaning US buyers are paying up again, a signal that's historically preceded sustained moves higher. CryptoQuant's CEO is now calling the bear market over, pointing to a profitability metric that matched the 2023 recovery pattern exactly.
The more interesting story is what's happening to miners. Bitcoin mining difficulty dropped 1.31% to 125.81 trillion, the tenth downward adjustment of 2026. Difficulty is down 15.1% year to date. Hashprice climbed 20% in four days. For miners still online, that's a real margin improvement — easier block work, more BTC per unit of hashrate.
And the equities are responding. Canaan, American Bitcoin and Cango surged between 41% and 67%. What's remarkable is that public miners have been selling BTC aggressively — MARA moved 23,093 coins in the first half of the year, RIOT dumped 9,665, together generating roughly $2.3 billion. Most of that funded AI and HPC pivots, not mining expansion. Public miners have plowed about $15 into AI data centers for every $1 of AI revenue they've booked. IREN's stock actually fell 8% today because the AI transition is eating margins.
So you've got a split market. Bitcoin exposure is finally paying off again for the pure-play miners. The AI-pivot miners are being punished for the capex. And underneath it all, difficulty relief is quietly rewarding whoever kept their rigs plugged in.
Meta released open weights for Muse Glimmer, a 30-billion-parameter model that fits under 20 GB and runs on a single consumer GPU. Apache 2.0 license. Optimized by Nvidia for their hardware. Aimed squarely at local, on-device agentic workflows — planning, tool calls, self-checks, failure recovery, coding tasks. It supports over 100 languages. And Meta pledged to open the weights for Muse Spark 1.2 soon, which adds parallel sub-agents, worktree isolation, and a crash-safe event log.
Zuckerberg paired the release with an essay titled "The Future Is for Everyone," arguing against a handful of institutions controlling advanced AI. Set aside the rhetoric — the strategic logic is straightforward. Meta doesn't sell API tokens. It sells attention on Instagram and WhatsApp. Every dollar OpenAI and Anthropic can charge developers is a dollar Meta would rather see burned to zero. Commoditize your complement. Open weights do exactly that.
There's a real technical wrinkle worth noting. Meta calls these open weight, not open source. The training data is not disclosed. That's a meaningful distinction the Open Source Initiative has been hammering on, and it's the same critique that dogged Llama 4 earlier this year when Meta got caught submitting a chat-tuned experimental variant of Maverick to LMArena leaderboards while shipping a different model to the public. Benchmarks got gamed. Trust took a hit.
Still, a 30-billion-parameter model that actually runs on your laptop and does agentic work reasonably well is a big deal. It's the size range where local inference stops being a toy and starts being useful. If you're building anything agentic that touches sensitive data — Bitcoin keys, financial workflows, private customer records — local models with real capability change the risk calculus completely. You don't have to pipe everything to a cloud provider that can subpoena, log, or train on your traffic.
Salesforce and Anthropic launched Claudeforce, and the framing is aggressive: Salesforce is telling customers they may never need to open the Salesforce app again. The centerpiece is a plugin with 37 prebuilt sales skills — meeting prep, deal health, pipeline reviews — that runs inside Claude and reasons over live CRM data. Claude also becomes the default model in Slack, powering Slackbot and Claude Code. Open beta starts in September.
The architecture matters. Actions route through MCP servers that enforce existing Salesforce permissions, so an agent can't do anything the user couldn't already do. Claude is the first LLM provider fully integrated inside Salesforce's Trust Boundary. And Claude can spin up a custom HTML dashboard on the fly — what the article calls a "vibe-coded" command center — while still pulling governed data from the backend.
Salesforce's own Agentic Enterprise Index puts hard numbers behind the shift. The average enterprise now runs 13 active AI agents, up from 5 a year ago — a 2.6x jump. Time to first agent deployment fell 53%, down to about two days. And roughly 70% of customer service sessions are being resolved fully autonomously, with escalation rates holding steady even at higher volume. Retailers who deployed agents last holiday season saw 4x higher sales growth than non-deployers.
The pattern here is bigger than Salesforce. The interface is dissolving. You stop opening apps and start talking to a model that opens apps for you. Slackbot alone logged 8.1 million hours of annualized productivity gains inside Salesforce, more than doubling quarter over quarter. Whether that number survives scrutiny or not, the direction is clear: the SaaS UI is becoming a headless API that an agent consumes on your behalf. Which means the moat isn't the interface anymore — it's the data, the permissions graph, and the governance layer. Everything else is a plugin.
The US Treasury finally proposed gatekeeping rules under the GENIUS Act this week. Here's the timeline that matters. January 18, 2027: stablecoin issuers must hold a federal or state license. July 18, 2028: exchanges, wallets, and DeFi interfaces are banned from offering unlicensed stablecoins to US persons. Public comment runs through October 19.
Foreign issuers get a path in, but it's narrow. They have to demonstrate ability to comply with US lawful orders and reciprocal arrangements. Intermediaries — the exchanges and wallets — are the gatekeepers, responsible for verifying that foreign issuers actually meet the bar. There's a safe harbor if a platform reasonably believes its customers are outside the US and doesn't advertise inward, but any actual knowledge of a US recipient blows that up.
The market is already repricing. USDT's share of US exchange volume dropped from about 72% in January to 64% in August. USDC climbed from 18% to 26%. Tether has responded by launching USAT, a US-focused stablecoin designed to be GENIUS-compliant from day one, while it waits on a reciprocity determination for the original USDT that hasn't come through. Circle, meanwhile, put USDC on Chelsea's Premier League jersey — a sponsorship reportedly worth up to $88 million a year — while giving legacy app developers 95 days to migrate off old cross-chain routes.
The consolidation logic is obvious. Reserves must be 100% Treasury bills or insured deposits. Weekly regulator reports, monthly disclosures, annual audits for issuers over $50 billion market cap. Small offshore issuers are being written out of the US market by design. And on a parallel track, the SEC proposed Regulation Crypto Assets, a new exempt offering framework with a $75 million fundraising cap, tiered disclosures on Form 1-CRYPTO, and preemption of state registration. Public comment there has 54 days left and, notably, almost no major exchange or asset manager has filed a comment yet.
One last note worth flagging: the Clarity Act slipped again, now targeting September. Banks aren't waiting. Every month without settled market-structure rules quietly rewards the incumbents who can afford to build inside walled gardens while everyone else waits for permission.
A prediction worth stress-testing: by this time next year, the enterprise SaaS interface as we know it will be a compliance layer, not a product. The product is the agent that talks to it. Whoever controls the permissions graph wins. Whoever built a pretty dashboard loses.