Bitcoin is under pressure again, trading around 62,900 dollars after fresh U.S. strikes on Iran rattled global markets. Korea's Kospi dropped 9 percent overnight, oil is up, and yet crypto is holding up better than stocks or bonds. Meanwhile, spot Bitcoin ETFs just snapped an 8-week outflow streak with 197 million dollars in fresh inflows. Mistral is about to drop a new open-weight frontier model in Europe. Stanford's Biomni AI agent is now running research workflows for over 10,000 labs. And Ledger's security team just cracked open flaws in both Trezor's new secure chip and Tangem's cards. Let's get into it.
Mistral is having a moment. The Paris-based lab is shipping a new open-weight model this July, with early access rolling out to research, government, and industry partners before a broader release later in the summer. They're calling it a new family of models, and the framing is pointed: this is Europe's shot at closing the frontier gap with OpenAI and Anthropic on open-weight terms.
Why does this matter beyond national pride? Two reasons. First, the timing. The EU AI Act's general-purpose provisions kick in on August 2nd, and Mistral has already signed the Code of Practice. For European enterprises, that means an open-weight model they can download, inspect, and run on their own infrastructure, under EU jurisdiction, right as the compliance clock starts ticking. That's a very different pitch from calling an API in Virginia.
Second, Mistral is putting real money behind the infrastructure. There's a 4 billion euro data center buildout across France and Sweden, including a hydropower-backed facility in Borlainge. They acquired Koyeb to build what they're calling a true AI cloud. And they just raised about 830 million dollars in debt financing for another data center near Paris loaded with Nvidia GPUs.
And it's not just language models. Mistral quietly launched Robostral Navigate, an 8-billion parameter robotics navigation model. Hardware-agnostic, works off a single RGB camera, uses online reinforcement learning to adapt. They're already talking to Airbus and BMW. Then last week they open-sourced Leanstral 1.5, a 119 billion parameter mixture-of-experts model focused on formal verification. It hit a perfect score on miniF2F, solved 587 of 672 PutnamBench problems, and reportedly used about one-seventh the compute of Claude Opus 4 to do it. In one test, they pointed it at 57 open-source Rust repos, found 47 property violations, and confirmed 11 real bugs — 5 of them previously unreported.
So the bet is coherent: sovereign infrastructure, open weights, formal verification, physical AI. If you're a European enterprise weighing whether to lock into a closed U.S. API contract right now, Mistral just handed you a reason to wait a few weeks.
There's a quiet shift happening in biomedical research, and it's worth paying attention to because it's already deployed, not hypothetical.
Stanford's Biomni project just published in Science. It's a general-purpose biomedical AI agent that behaves like a co-scientist. You give it a plain-language question. It reads the literature, picks the right tools from a library of 150, pulls from 59 databases, writes the analysis code, runs it, and generates hypotheses. On tests like single-cell RNA-seq annotation, rare disease diagnosis, and GWAS causal gene detection, it matched human experts — but dramatically faster. They pointed it at wearable sensor data and it spat out plausible hypotheses in about 40 minutes. That's work that would take a graduate student most of a week.
And here's the number that matters: over 10,000 labs are already using it. It's open-source. This isn't a demo — it's the most widely deployed AI co-scientist in biomedicine, and it happened without a lot of noise.
Meanwhile in China, MGI Tech and Shanghai AI Lab unveiled two related pieces: ProtoPilot and BioLab Bench. ProtoPilot is a multi-agent system that spans the entire experimental lifecycle — design a protocol, generate the code, execute on a device, get wet-lab feedback, learn from failures, regenerate. In their reported metrics, ProtoPilot hit 52.38 percent on complex protocol tasks versus 54 percent for human experts and 43.5 percent for GPT-5.6-sol. BioLab Bench is the evaluation framework — it tests whether an AI agent can actually run real experiments on automated lab platforms, not just write plausible-sounding instructions. The goal is a 24/7 unattended intelligent lab.
And then there's Aureka's OpenDDE — an open-source, all-atom biomolecular foundation model for drug discovery. 655 million parameters, trained on roughly 414,000 GPU-hours. It's aimed at antibody-antigen co-folding and eventually de novo drug design. Top-1 antibody-antigen benchmark scores in the 51 to 70 percent range, jumping to 66 to 82 percent under oracle selection.
The pattern here is that AI agents are moving from chatbot demos to running real scientific workflows, on real instruments, at scale. The friction between hypothesis and experiment is collapsing. Whether that translates into faster drugs or faster papers is the next question — but the tools are already in production.
Bitcoin dropped below 63,000 dollars overnight after the fourth round of U.S. strikes on Iran. Oil jumped, yields rose, the dollar strengthened, and equity futures fell. Korea's Kospi lost 9.2 percent. About 253 million dollars in leveraged crypto positions were liquidated — significant, but only about a sixth of what we saw at the worst points in the past 30 days. Bitcoin is holding around 62,900 to 63,800 depending on the hour, and the notable thing is that it's holding better than almost anything else on the risk board. Prediction markets are pricing just a 3 percent chance the Strait of Hormuz traffic normalizes by August. So this isn't going away this week.
But zoom out. For the week ending July 11th, U.S. spot Bitcoin ETFs pulled in 197 million dollars in net inflows across 13 products. That ended an eight-week outflow streak that had drained more than 8 billion dollars from the sector. BlackRock's IBIT did the heavy lifting, as usual — around 87 million dollars on July 10th alone. IBIT's cumulative inflows since January 2024 now sit near 60 billion dollars. Total net assets across spot Bitcoin ETFs are around 77 billion, roughly 6 percent of Bitcoin's market cap.
So you have this split picture. Short-term, geopolitics is dragging price. Longer-term, regulated demand quietly turned positive again. Standard Chartered analysts note that Michael Saylor's messaging around Strategy's pivot is muddying the waters — Strategy just raised 467 million dollars in cash by selling MSTR shares, leaving their 843,775 Bitcoin stack completely untouched. Their U.S. dollar reserve is now at 3 billion. The market read it as ambiguity about whether the flywheel still spins the same way.
One other data point worth flagging: a Bitcoin whale moved 188 million dollars after seven years of dormancy. Old coins waking up is not automatically bearish, but combined with the war headlines it's a reason short-term holders are jumpy. CryptoQuant data shows holders in the 1-to-6-month bucket are roughly 15 percent underwater. If price pushes back toward 71,000 to 77,500, expect that cohort to sell into strength.
Resistance is 65,000, then a supply wall between 71,000 and 77,500. Support is 60,000. That's the range that matters this week.
If you self-custody Bitcoin, this week had some uncomfortable news. Ledger's Donjon security team — which is genuinely one of the best hardware wallet audit groups in the world — dropped two separate findings.
First, they found a vulnerability in the TROPIC01 secure element chip used in Trezor's new Safe 7. Trezor is being upfront about it and says user funds remain safe because of multi-layer security, but the disclosure is real, and it's a reminder that even purpose-built secure elements can have flaws.
Second, and more dramatic: Donjon demonstrated a laser fault-injection attack on Tangem wallet cards. Using a nanosecond laser pulse aimed at the Samsung S3D232A secure element, they can bypass the password recovery check and reset the card's password without knowing the original. The chip has EAL6-plus certification, which is supposed to be top-tier. Didn't matter — the vulnerability is in the recovery logic, not the crypto.
Here's the practical framing. The attack requires physical possession of the card, about 250,000 dollars of lab equipment, roughly two hours of work per card, and it visibly damages the hardware. So no, someone is not draining your Tangem remotely. If your card is in your pocket, you're fine. But — and this is the part that matters — Tangem cards have no firmware update mechanism. That was originally a security feature: no remote code changes possible. But it also means this flaw cannot be patched. Ever. On any card currently in circulation.
For everyday users with modest holdings, the risk is minimal. For high-net-worth or institutional custody, an unpatchable vulnerability in your storage device is a different conversation. If you lose a Tangem card, treat it as a security event — rotate funds to a new wallet.
One more thing worth noting alongside this. After the MiCA deadline in Europe, Binance said about 70 percent of EU users withdrawing funds sent them to self-custody wallets rather than to other MiCA-compliant exchanges. That's a striking number. Regulators wanted people moving to supervised venues. Users largely voted with their private keys instead. Self-custody is winning the trust battle even when the tools have known flaws — which is exactly why audits like Donjon's matter.
Here's what's actually interesting about today: Europe is quietly executing on sovereign AI infrastructure, AI agents are already writing code and running experiments in thousands of real labs, Bitcoin absorbed a fourth round of Middle East strikes better than gold or equities, and 70 percent of Europeans leaving Binance chose their own keys over another exchange. Four different domains, one direction — power moving down the stack, closer to the user. Ignore the noise. Watch that trend.