Bitcoin slipped below $63,000 as spot ETFs posted back-to-back outflows for the first time since late July. The SEC abruptly canceled its Reg Crypto meeting, punting the first major rulemaking of this cycle. Tether finally landed a clean KPMG audit, showing reserves exceeding liabilities by $6.8 billion. And Mistral just made its biggest strategic move yet — trying to turn Europe into a real AI compute power. Four stories worth your time today.
Mistral dropped something bigger than a model release yesterday. The French AI lab is pivoting from being a model maker to being AI infrastructure for Europe — and the ambition is enormous.
Three pieces to the announcement. First, Regional Endpoints are now generally available. Customers pick whether inference runs in Europe or the US, with data residency guarantees. Second, a Priority Tier in public preview — committed SLAs, 99.5% uptime, custom rate limits, priced at 1.75x standard. Third, and most interesting: Mistral is going to host third-party open models on its own infrastructure, starting with Z.ai's GLM-5.2 out of China. Pricing lands at $1.40 per million input tokens, $4.40 per million output.
Think about what that means. A European enterprise can now run a Chinese open model, on French infrastructure, under EU data rules, through a single API. That's the AWS Bedrock playbook, but with sovereignty as the pitch.
The real bet, though, is European Compute Units. Mistral is assembling an anchor coalition — ASML, Amadeus, Capgemini, Caisse des Dépôts, CMA CGM, and Microsoft as a multi-billion-dollar customer — to commit to multi-year compute purchases. Current European capacity sits under 200 megawatts. The target: 200 megawatts by end of 2027, and up to 1 gigawatt by 2030.
Here's the strategic read. Mistral is quietly admitting that individual models may not be where the durable value sits. Infrastructure is. If you own the compute network that European regulated industries are legally comfortable running on, you capture demand whether or not customers use your models. It's a defensive moat dressed up as sovereignty. And honestly? Given how fast open weights are catching up to closed labs, this might be the smarter bet.
Three stories this week make it clear AI in education is moving from demos to actual deployments — with real students, real tuition, real contracts.
Drew University in New Jersey is opening a K-8 school this fall powered by 2HR Learning's TimeBack platform. Small cohort — 10 to 12 founding families at $30,000 a year. The pitch: two focused hours of adaptive tutoring for reading, writing, and math, then the rest of the day on projects and practical skills. A university campus running an AI-driven elementary school is a genuinely new format.
Andrew Ng's Kira Learning is going the other direction — putting AI agents inside existing K-12 classrooms. Grading, lesson planning, analyzing which students are struggling in real time. Ng's framing is that AI lets teachers become better teachers, especially in subjects like computer science where most teachers lack deep background.
And in Malaysia, LAPA Labs and HELP University reported 100% module completion in an adaptive AI curriculum pilot, and just signed an MOU to expand into corporate training targeting 500-plus employees in phase one.
Here's what's actually interesting. These aren't chatbots stapled onto worksheets. They're end-to-end workflows — the AI decides sequencing, tests recall, grades output, and flags struggling students. The bottleneck isn't the technology anymore. It's whether coaches and teachers can integrate this into a coherent day, and whether learning outcomes hold up under independent measurement. Nobody has produced that evidence yet. Kira has no published outcome data. 2HR Learning cites its own internal MAP Growth numbers. Skepticism is warranted.
But the market is voting. When families pay $30,000 for AI-led instruction, and universities lend their brands to it, the traditional edtech stack has a problem.
The Q2 numbers are in for public Bitcoin miners, and the transformation is happening faster than most people realize. Public miners collectively shed 21% of their hashrate — excluding Bitdeer, which pulled the other direction. The story is a wholesale reallocation of power capacity from Bitcoin mining to AI and high-performance computing.
Riot Platforms is the poster child. They just signed a 20-year, 191-megawatt lease at their Rockdale campus with what they're calling a frontier AI lab. Initial contract revenue: roughly $9.1 billion, potentially $16.1 billion with extensions. Combined with their AMD deal, Riot now has 241 megawatts under contract for AI tenants. Estimated cumulative net operating income over the base term: $7.3 to $8.2 billion.
Core Scientific's colocation revenue hit $136.7 million last quarter — 83% of sales — versus $27.5 million from mining. TeraWulf: 71% of revenue now from HPC leasing. MARA swung to a $611 million net loss on Bitcoin's price drop, even while producing 2,422 BTC, and is now targeting at least two AI leases by year-end plus a 2-gigawatt Texas site.
Bitdeer is the outlier — quintupled mining output to 2,694 BTC in Q2, hashrate up 389% to 69.5 exahash. But even they signed a 16-year, $4.7 billion AI lease in Norway.
The network hashrate itself dropped from 1,071 exahash to 957 exahash quarter-over-quarter. That's a 10.6% network decline, driven by economics, not any single event. Post-halving mining margins plus insane AI compute demand equals miners becoming, effectively, power landlords with hash rate as a side business.
For Bitcoin, this is a mixed signal. Difficulty adjusting down helps remaining miners. But the biggest infrastructure operators in the ecosystem are quietly telling you where they think the durable cash flows live — and it's not in block subsidies.
Two market stories worth connecting. Spot Bitcoin ETFs saw their first back-to-back daily outflows since late July this week, as Bitcoin slipped below $63,000. That comes right after a strong week — $853 million of inflows led by BlackRock's IBIT taking roughly 80% of the total. Year-to-date, though, Bitcoin ETFs are still net negative by around $4.4 billion.
The institutional picture is mixed but interesting. JPMorgan disclosed a 25% increase in its Bitcoin ETF position and more than quadrupled its Ether ETF position in Q2. Morgan Stanley's Bitcoin Trust has drawn $371 million in share contributions since April launch without a single day of outflows. Israel's largest bank, Bank Leumi, will offer Bitcoin, Ether, and Solana trading through its investment app starting early 2027, partnering with Galaxy.
But the regulatory side just took a hit. The SEC abruptly canceled Friday's meeting to propose Reg Crypto — the first major crypto rule of this administration. No new date. They're also delaying the innovation exemption for tokenization amid Wall Street and White House concerns. The cancellation followed the Senate leaving for recess without voting on the CLARITY Act. The CFTC is now stepping in with its own Innovation Advisory Committee meeting on August 20.
And MSCI floated a proposal that could exclude Strategy and Metaplanet from major indices as non-operating companies. Strategy fired back that index providers should measure markets, not dictate what public companies are allowed to own. That fight matters — passive index flows are a meaningful chunk of the Bitcoin-treasury-company thesis.
Bitcoin looks caught between weakening spot demand, sticky Treasury yields, and a regulatory calendar that just slipped again. The S&P 500 hitting 7,798 while Bitcoin sits at $63,000 tells you where the marginal risk appetite is going right now.
Notice the pattern across today's stories. Mistral is buying compute. Riot is leasing compute. Miners are pivoting to compute. The scarce resource of this cycle isn't models, isn't tokens, isn't even capital — it's power delivered to a rack. Whoever controls that in five years controls a lot more than they realize today.