Bitcoin clawed back above 62,000 dollars on a weak US jobs report that pretty much killed any remaining talk of a Fed hike. Spot Bitcoin ETFs finally snapped a brutal 10-day outflow streak with 221 million dollars in fresh inflows. Meanwhile, Mark Zuckerberg told Meta staff that AI agents aren't advancing as fast as he hoped, even as his AI chief claims their next model has caught up to GPT-5.5. And in Europe, MiCA's grandfathering window slammed shut on July 1, reshaping who can legally serve crypto customers across the EU. Let's get into it.
Meta had one of the more revealing weeks in AI this year, and it's a study in contradictions. At an internal town hall, Mark Zuckerberg admitted to staff that AI agents haven't progressed as quickly as he'd hoped. He said the trajectory over the past four months hasn't accelerated the way he expected, and that his aggressive restructuring, which included laying off around 8,000 corporate staff and reassigning 7,000 more into AI groups like Agent Transformation, hasn't yet paid off. His words, not mine. He said the upside hadn't materialized, though he expects improvements in the next three to six months.
Now, at the same town hall, Meta's AI chief Alexandr Wang was singing a very different tune. Wang claimed that Meta's upcoming model, codenamed Watermelon, has already caught up with OpenAI's GPT-5.5 on internal benchmarks. Watermelon is the successor to Muse Spark, codenamed Avocado, and it's reportedly being trained on dramatically more compute. Wang also promised a Claude Opus-level coding model, quote, pretty soon.
So which is it? Behind schedule or catching the leaders? Probably both, and that's the point. Meta is spending up to 145 billion dollars on AI infrastructure this year. That's not a typo. Part of that broader 700-billion-plus big tech AI capex wave. When you're spending that kind of money, benchmark parity on an unreleased model is table stakes. The real question is whether any of this translates into agents that actually work in the wild.
And here's the interesting bit for anyone watching markets: memory and semiconductor stocks that dominated 2026 are losing momentum. Some of that capital appears to be rotating back into Bitcoin. If the AI trade cools even a little, and Meta's admission is a small crack in that narrative, the second-order effects on crypto liquidity could be meaningful. Zuckerberg basically told his own company the honeymoon isn't happening on schedule. Markets noticed.
While Zuckerberg is managing expectations on consumer AI agents, the enterprise side is quietly shipping. Two announcements this week worth your attention.
First, Microsoft made Service Agent generally available inside Microsoft 365 Copilot. This isn't a chatbot. It plugs Dynamics 365 customer service workflows into Copilot with 70-plus new MCP tools and 20-plus product enhancements. Agents can pull case data from Dataverse, SharePoint, and Microsoft 365, draft customer communications, update cases, and take actions inside a single chat window without tab-hopping across five systems. The key word there is actions. Not summarize, not suggest, actually do the thing.
Second, Genesys acquired Pinkfish. If you haven't heard of Pinkfish, they're a workflow automation shop with 500-plus pre-built connectors and 25,000 MCP tools linking AI agents to CRM, ERP, ITSM, HR, order management, and billing systems. And that's the whole story right there. The thing that's been killing enterprise AI agents isn't reasoning. It's integration. An agent that can think brilliantly but can't touch your order system is a very expensive chatbot. Genesys is buying its way out of that gap.
And over in Europe, Sopra Steria rolled out NiCE CXone and Copilot for Agents across 2,000 employees in France, Poland, and India in under three months. 90% of calls answered within 20 seconds. Boring, unglamorous, but this is what AI actually looks like when it's working.
The pattern here matters. The winners in agentic AI probably aren't going to be the labs shipping the biggest foundation models. They're going to be the boring middleware companies that connect models to the fifty enterprise systems every real company actually runs on. MCP, Model Context Protocol, is quietly becoming the standard that makes this whole thing work. Watch that space.
Bitcoin traded up to just over 62,000 dollars, gaining nearly 4% on the day after US June payrolls came in at 57,000 versus an expected 110,000. Prior months were revised down by a combined 74,000. Unemployment ticked up to 4.2%. Traders read this as the Fed rate cut catalyst they've been waiting for. Bearish traders got liquidated for 281 million dollars in 24 hours, roughly double the long liquidations, in a classic short squeeze.
But here's the more interesting story underneath the price. In the two weeks leading into this rebound, Bitcoin whales bought 16.7 billion dollars of BTC. That's roughly 270,000 coins absorbed by large holders. At the same time, US spot Bitcoin ETFs bled a record 4 billion dollars in June, their worst month ever. Institutional retail-facing demand cratered while whales stepped in with both hands. That divergence has historically shown up near cycle bottoms, not tops.
Then yesterday, ETFs finally posted 221 million dollars of inflows, their best day in two months, and notably not driven by BlackRock's IBIT this time. So a broader base of buyers waking up.
On-chain, though, some caution. Whales sent 49,000 BTC to exchanges during the selloff, one of the largest daily inflows of the year, which usually signals selling pressure. Options markets aren't fully buying the bounce either, according to Coindesk, with skew still leaning defensive.
And Bitwise had a spicy take on Strategy. Their STRC product got hammered this week, and Matt Hougan basically said, look, offering high yields and low volatility was always a weird fit for a Bitcoin proxy since Bitcoin offers neither. Bitwise thinks Strategy will become, quote, less important in the Bitcoin story going forward, with new institutional buyers stepping in to replace them. If that's right, this whole cycle's leverage structure gets a lot healthier on the other side of it. The 14% Q2 drop happened alongside the first stablecoin market contraction since 2023, total supply slipping to 312 billion dollars. Liquidity was genuinely tight. Now watch if it comes back.
MiCA's grandfathering window officially closed on July 1. If you're a crypto service provider operating in the EU without a MiCA license, you're winding down. This is separating the survivors from everyone else, and the survivors this week are Bitcoin-first companies.
Estonia granted its first standalone MiCA CASP license to Lightspark, paired with an electronic money institution license from the country's Financial Supervisory Authority. That dual authorization is huge. It means Lightspark can passport payment and crypto services across all 29 EU and EEA markets from a single Estonian base. Stablecoin on and off ramps, cross-border payouts, and Lightning-based settlement, all from one license. The MiCA capital requirement starts at 150,000 euros for higher-risk activities, so it's not free, but compared to licensing individually across 30 jurisdictions, this is a shortcut worth taking.
Even more interesting: Wavespace launched the wavecard, a MiCA-compliant, self-custodial Bitcoin debit card for the Eurozone. Read that phrase again. Self-custodial. Bitcoin. Debit card. MiCA-compliant. Those things weren't supposed to coexist. The card uses Nostr Wallet Connect, specifically NIP-47, to draw sats directly from your own Lightning node at every Visa purchase. No pre-loading into a Wavespace custodial wallet. You grant a scoped connection key with a spending limit that you can revoke at any time. Wavespace says 70% of platform payments already run over Lightning. They're integrated with Lightspark, exploring the ARK protocol, and looking at US expansion. They're currently fundraising after bootstrapping with initial support from Relai.
Meanwhile in Kenya, Tando is doing something even more interesting from an adoption standpoint. Their app lets M-Pesa users receive Bitcoin over Lightning that gets converted to Kenyan shillings in real time and deposited into the merchant's M-Pesa account. No wallet needed on the receiver's end. Around 40 million M-Pesa users are effectively Lightning-addressable via their phone numbers. Over 100 daily transactions by mid-2025, zero extra transaction fees. This is the spend not sell movement in Africa, and it's the clearest example anywhere of Bitcoin as actual medium of exchange rather than a treasury asset.
Binance, by contrast, withdrew its MiCA application days before the deadline. Their Europe head is now arguing MiCA should be judged by who it licenses, not who it excludes. Which is one way to spin it. The other way is that the Bitcoin-native players just quietly built the infrastructure and got the licenses while the big exchanges were still negotiating.
One prediction to close on. The gap between what AI labs promise and what actually ships is starting to bite even Meta, spending 145 billion dollars a year. Bitcoin, meanwhile, just had its worst institutional month ever and whales absorbed all of it. Watch which of those two narratives ages better through the second half of the year.