Bitcoin is hovering just above $64,000 as markets brace for a Fed decision that most expect to be a hold, but a real minority, including Citadel Securities and UBS, is warning about a surprise hike. Spot Bitcoin ETFs have now bled $526 million over four straight sessions. Meanwhile, Microsoft cut a multibillion-dollar deal with Mistral to build out European AI infrastructure, and unveiled its own agentic cybersecurity stack with a compact in-house model. Core Scientific doubled revenue on the back of AI colocation and added 301 more Bitcoin to its treasury. And a Mexican billionaire is anchoring a $40 million bet on a permanent-capital Bitcoin holding company. Let's get into it.
So here's the setup for today's Fed meeting, and why it actually matters more than the usual ritual. Bitcoin dipped as low as $62,900 in the past 24 hours before recovering to around $64,000. Inflation is running at 4.1%, which keeps a rate hike genuinely on the table. Now the base case among most analysts is a hold. But Citadel Securities and UBS are flagging a real chance of a surprise hike, and Citadel's macro team has an interesting argument: this hike wouldn't be about the data at all. It would be about Fed Chair Kevin Warsh taking his best shot at making a surprise actually land, before the market stops being surprised by anything he does. Warsh has broken with a 30-year playbook by essentially eliminating forward guidance. No hand-holding, no telegraphed moves. That's a huge shift, and it leaves Bitcoin unusually exposed. Because for the last decade, crypto has traded like a leveraged bet on Fed liquidity. Dovish signals pump, hawkish signals dump. One analyst put it well: anything remotely dovish could be good for Bitcoin. But if Warsh actually hikes, or even just refuses to hint at cuts, Bitcoin's downside is wide open, especially with ETFs already in outflow mode. What's genuinely interesting is that Bitcoin has held up better than AI-driven tech stocks through the recent volatility. When SK Hynix dropped 17% despite profits rising 557%, and that contagion spread through Asia, Bitcoin only slipped about 1%. Korea's chip crash barely touched crypto. That's not nothing. It suggests Bitcoin's correlation profile is quietly shifting. It's less the pure risk-on proxy it was two years ago, and more its own thing. Watch the reaction function today. If Bitcoin holds $63,000 through a hawkish surprise, that's a real tell about who's actually holding the float right now.
Microsoft just signed a multibillion-dollar deal with Mistral, and this one is more interesting than the usual big-tech AI announcement, because it's really about geography. Under the agreement, Azure customers can now run workloads in Mistral's data centers in France. Mistral's Medium 3.5 and OCR 4 models are being folded into Microsoft Foundry and Copilot Studio. And critically, the setup supports Azure Local, which means enterprises can run these models fully offline, on their own hardware, for regulated or sensitive workloads. Mistral is putting €4 billion into European data centers, with a facility open near Paris, another under construction in Sweden, and a target of 1 gigawatt of compute capacity by 2030. The pitch is European sovereignty. Reduce dependence on US-controlled infrastructure. Give regulated industries a real alternative. Now, is it actually sovereign? Not entirely. The GPUs are still Nvidia. The hyperscaler partner is still American. But the data residency and operational control angle is real, and it clearly resonates with European buyers. Microsoft didn't take new equity in Mistral this round. But there are unconfirmed reports of an upcoming €3 billion funding round at a €20 billion valuation, with Samsung reportedly considering a €1 billion check. That would be roughly 70% higher than Mistral's €11.7 billion valuation from the ASML-led round just months ago. The bigger story here: the era of one global AI stack is ending. Export restrictions from Washington, EU sovereignty demands, and regulated-industry compliance requirements are fracturing the market. What Microsoft is doing is smart. Rather than fight that fragmentation, they're partnering their way into every jurisdiction that matters. Mistral gets scale and distribution. Microsoft gets a sovereignty story to sell to European regulators and banks. The losers are pure-play frontier labs that don't have a local partner in every region that cares.
Microsoft also dropped something else worth paying attention to: Project Perception, an agentic cybersecurity platform, and a new in-house model called MAI-Cyber-1-Flash. The architecture is what's interesting. Three types of agents work together. Red team agents look for attack paths before exploitation. Blue team agents analyze context and assess real risk. Green team agents actually implement fixes, patches, WAF blocks, remediation code. It's a closed loop that runs continuously. The economics are the real story though. MAI-Cyber-1-Flash is a small, specialized security model that handles about 90% of typical cybersecurity tasks. For the harder 10%, the system escalates to GPT-5.4, which Microsoft explicitly chose over the more expensive GPT-5.6 for cost reasons. The combined stack scored around 96% on the CyberGym benchmark, beating frontier-only setups, and cut costs roughly in half compared to Microsoft's previous configuration. This is the shift that's been coming for a while. The competitive edge in enterprise AI isn't the biggest model anymore. It's the orchestration. It's the routing layer that knows when to use a cheap specialized model and when to spend on the expensive one. For always-on workloads like security monitoring, where you're running inference 24/7 across millions of endpoints, the cost curve of frontier-only stacks just doesn't work. There's a broader lesson here for anyone building AI products. If your architecture assumes GPT-5.6 for everything, you're going to get out-priced by someone running a well-designed multi-model harness. Public preview of MAI-Cyber-1-Flash starts next week. Project Perception's public preview kicks off August 3rd. Expect competitors to respond fast, because Microsoft just showed the playbook.
Two Bitcoin-adjacent stories worth connecting. First, Orange Juice Holdings. Mexican billionaire Ricardo Salinas, who reportedly holds about 80% of his liquid portfolio in Bitcoin, anchored a $40 million seed round for a permanent-capital holding company co-founded by macro strategist Lyn Alden. The pitch is Berkshire Hathaway meets the Bitcoin standard. Acquire profitable US businesses at low multiples, hold them permanently with no forced exit timeline, and funnel excess cash flow into a Bitcoin treasury. Salinas says hundreds of investors and operators have reached out since the July 15 announcement. What's notable is that this isn't a crypto-native thesis. It's traditional business capital, people who understand cash flows and operating companies, being drawn to a structure that solves a real PE problem: the tyranny of the fund lifecycle. Permanent capital plus a hard-money treasury. If Orange Juice actually deploys well, expect copycats. Second, Core Scientific's Q2. Revenue more than doubled year-over-year to $164 million, driven almost entirely by AI colocation. They also added 301 Bitcoin to bring holdings to 848 BTC, worth around $54 million. And they announced a partnership with AMD for up to 2.5 gigawatts of data center capacity, with $14 billion in base contracted revenue on the initial 530 megawatts. Total contracted future payments now exceed $24 billion. But here's the caveat. GAAP net loss for the quarter was $1.16 billion, mostly non-cash. And CryptoSlate raised a genuine concern: Bitcoin miners are pouring billions into AI right at the moment mining profitability is near historic lows. If the AI capex cycle cools while their mining fleets have already been repurposed or sold, some of these pivots could look premature within a year. Core Scientific has $6 billion in contracted backlog to insulate them. Smaller miners chasing the same trade may not.
One prediction for the rest of the week: whatever the Fed does today matters less than what Warsh doesn't say afterward. The absence of guidance is the new signal.