Google DeepMind drops a new Gemini family with a robotics play that finally moves past tabletop demos. Lightning infrastructure just had its worst week in years, with Boltz, ZEUS, AQUA, and BTCPay all hit inside a few days. Bitcoin is stuck below 65,000 heading into a big CPI print. And Riot Platforms just signed a 9.1 billion dollar compute deal with Anthropic, which tells you exactly where miner economics are heading. Let's get into it.
Google DeepMind rolled out a big update this week, and the headline isn't another chatbot benchmark. It's Gemini Robotics 2, and for the first time it feels like an AI model that actually controls a body.
The setup is three parts. There's a vision-language-action model that translates what a robot sees and hears into full-body motion. There's Gemini Robotics ER 2, which handles the planning and reasoning. And there's an on-device version that runs locally on the hardware. The interesting bit is that one model checkpoint can control completely different robot bodies. A humanoid with five-fingered hands, a two-armed Franka platform with a gripper, whatever. No retraining per hardware variant. Adaptation to new hardware takes hours and fewer than 200 examples.
Apptronik's Apollo 2 is the flagship demo, and the numbers are honest in a way company demos usually aren't. Success rates on table pickups around 68%. Floor pickups drop to 45%. Shelf work climbs to 76%. Fine dexterity is still rough. Screwing in a bulb, 36%. Sealing a zip-top bag, 40%. Unscrewing a bulb, 92%. So robots are great at destruction and mediocre at assembly, which honestly tracks.
On the software side, three new Gemini models. 3.6 Flash for high-quality agent actions, 3.5 Flash-Lite for cheap repetitive loops, and 3.5 Flash Cyber for security workflows. Pricing is aggressive, roughly 2 dollars per million input tokens and 10 dollars per million output. That's cheap enough that small teams can actually run agent loops in production without watching the meter.
And there's a leadership shuffle worth noting. Demis Hassabis moves up to Chair of DeepMind and Chief Scientist of Alphabet. Koray Kavukcuoglu takes over day-to-day model development. The Gemini app is at 950 million monthly active users, Gemma has crossed 900 million downloads. Whatever you think of Google's product design, the distribution is real.
The honest read here: whole-body robotic control from a single generalist model is a genuine step change. Whether Apollo 2 can hold a shift on a Mercedes-Benz line without a human babysitter is the question that matters, and we don't have that data yet.
While Google is teaching robots to walk, the rest of the industry is teaching agents to ship code without supervision. And this week made it clear that autonomous software engineering has moved from demo to deployment.
Meta launched Muse Code with Muse Spark 1.2, a terminal coding agent built around a persistent multi-agent system. There's a main agent and long-running background subagents that decide on their own when to check in. Everything gets logged to a local event log, so if a machine crashes mid-run, it replays exactly where it left off. They used it internally to iterate on GPU kernel optimization, over 1,000 tool calls in a single 24-hour run.
AWS shipped Kiro Crew, an open-source orchestration platform for running long-lived agent teams across your repos, your CI, your issue tracker. Self-hosted, persistent memory, approval workflows, signed audit logs. The pitch is that enterprises don't want their code and credentials leaving the perimeter, so give them the orchestration layer and let them keep the keys.
And Goldman Sachs is running Devin in production alongside 12,000 human engineers. They're reporting 3 to 4 times the productivity of prior AI tools, and they're now scaling to Claude for trades, transactions, and client onboarding. The estimate floating around is 200,000 US banking jobs at risk, disproportionately junior developers.
That last bit is the uncomfortable part. If you kill the junior rung, where do senior engineers come from in 10 years? Nobody has an answer, and Goldman's CTO isn't paid to worry about it.
On the security side, GitLab now supports confidential AI via Privatemode, running inference inside AMD and Intel trusted execution environments with remote attestation. Your code goes into an encrypted enclave, gets processed, comes back. Neither the model operator nor the cloud provider can see it. That's the pattern regulated industries have been waiting for.
The throughline across all of this: agents are getting persistent memory, orchestration, and enterprise-grade governance. The demo phase is over. The procurement phase has begun.
Now for the bad week in Bitcoin infrastructure. Really bad.
Inside a 72-hour window, three Lightning-adjacent services went down. Boltz, the swap provider that a lot of wallets quietly depend on, shut off all swap services on August 3rd after what they described as AI-powered automated attacks that had been probing their code for months. No customer funds lost because Boltz uses hash time-locked contracts and doesn't custody, but wallets like Aqua, Bull Bitcoin, and Zeus all lost swap functionality overnight.
Then ZEUS pulled its own Lightning infrastructure offline after a separate cybersecurity incident. Same message: contained, no funds lost, but liquidity provider channels get terminated during downtime.
And then the ugly one. BTCPay Server had a critical vulnerability where LND macaroons, the credentials that authenticate Lightning nodes, persisted after updates. Attackers used old but still-valid credentials to drain merchant Lightning wallets. Foundation and hodlonaut both got hit. Channels force-closed, funds swept. BTCPay released version 2.4.2, is offering a bounty of up to 190,000 dollars for recovery, and is telling operators to either upgrade immediately or shut servers down.
Meanwhile, the network itself is fine. Stacker News snapshots this week show around 17,400 nodes, roughly 41,000 channels, and 4,900 BTC of capacity. A 5 BTC channel just opened between LNBiG and ACINQ. Bitfinex's primary Lightning node is sitting on 365 BTC across 606 channels. The rails work.
The problem isn't Lightning. The problem is the layer of infrastructure providers sitting on top of it, and specifically the concentration of dependencies. When one swap provider going down disables features across half a dozen wallets, that's a single point of failure dressed up as decentralization. Aqua, Bull Bitcoin, Blockstream, and Zeus are all now openly talking about redundancy and backup providers, which is exactly the right response, just about three years late.
And the AI-powered attack angle on Boltz deserves attention. If adversaries can run automated code-probing at scale against open-source Bitcoin infrastructure, every self-hosted operator needs to assume their attack surface just got a lot bigger.
Bitcoin is trading around 64,000 heading into Wednesday's CPI print, stuck between what Glassnode calls a 63,000 on-chain demand shelf and 69,000 holder resistance. The Strait of Hormuz relief trade fell apart after Trump demanded 50 years of Iranian compensation, oil bounced back to 89 dollars, and the whole risk complex went sideways.
But the more interesting signal is where mining money is flowing. Riot Platforms surged 20% in pre-market after announcing a 9.1 billion dollar, 20-year deal with Anthropic for 191 megawatts at its Rockdale, Texas campus. Keel just shut down its US Bitcoin mining operations entirely and is pivoting to AI and high-performance computing. Q2 revenue fell 50%.
This is the trade of 2026. Public miners with power contracts and cheap land are worth more as AI landlords than as hash producers. Nvidia is signing MOUs with six Wall Street firms to turn AI compute into a bankable asset class. FxPro's Alex Kuptsikevich put it bluntly: companies that gave Bitcoin institutional cover are now chasing AI data centers instead.
On the sovereign side, the numbers are worth knowing. Thirteen governments now hold about 26.8 billion dollars in Bitcoin, US in the lead, mostly through seizures. El Salvador quietly bought its 1 BTC per day, up to roughly 7,734 BTC total. Bhutan went the other direction, selling another 434 BTC for about 28 million to fund the Gelephu Mindfulness City. Their state mining pipeline has slowed, so they're drawing down without replenishment.
Bitwise data shows elite wallets holding over 10,000 BTC hitting a six-month high of 90. Bitcoin-backed lending is being called institutional-grade by Two Prime as public companies borrow against BTC instead of selling. BlackRock Canada launched a new ETF with a 3% Bitcoin allocation baked in.
And Strategy is now sitting on a 4.75 billion dollar cash cushion, having learned that traditional investors don't view its BTC stack as a dollar substitute. Trump Media's crypto arm posted a 361 million dollar loss and is restructuring its treasury approach after the Q2 damage.
The pattern is consistent. Bitcoin the asset keeps getting absorbed by institutions and states. Bitcoin the industry, particularly the mining side, is being partially eaten by AI. Both things can be true at the same time.
If you run a Lightning node, a BTCPay server, or anything self-hosted in this ecosystem, rotate your credentials this week. Not next quarter. This week. The attackers now have AI on their side, and the humans defending open-source infrastructure do not have the same headcount advantage they used to.