Bitcoin is coiled around $64,000 while global equities print fresh records. The Coldcard hardware wallet exploit keeps expanding, with Galaxy Research now counting at least 15 different attackers and roughly 1,596 BTC stolen from about 7,300 addresses. Elon Musk says Grok 4.6 lands next week, with 4.7 close behind and Grok 5 slated for year-end trained on SpaceX's full dataset. Bitcoin ETFs shook off a $265 million outflow day with $382 million of inflows over the next two sessions. And Arthur Hayes is back with his loudest call yet — an AI credit bubble that eventually cracks and paves Bitcoin's road to $1 million.
Let's start with xAI, because the cadence Musk is describing is genuinely unusual. Grok 4.5 dropped July 8. Grok 4.6 is expected around August 7 — a 1.5 trillion parameter model with upgraded fine-tuning and reinforcement learning. Grok 4.7 follows weeks later at 2.1 trillion parameters. And Grok 5, which is supposed to be trained on the entire SpaceX data footprint, is aimed at year-end. That's four major model releases in under six months from a single lab.
The infrastructure story behind it is arguably bigger than the models themselves. On SpaceX's earnings call yesterday, Musk said online AI compute is targeted to cross 2 gigawatts this year, 10 gigawatts by end of next year, and roughly 15 gigawatts by 2027. All of it running on Nvidia's Vera Rubin architecture, which he called the best AI computer on the market. And they're planning to extend compute into orbit through Starmind satellites carrying Rubin GPUs — data center grade AI in space, with the first birds launching next year.
Meanwhile, xAI is redeploying engineers off Starlink and Starship to accelerate Grok, and integrating the model into Cursor to compete directly with Codex and Copilot. The pitch has shifted from chatbot to autonomous agent — coding, real-time synthesis, execution.
Two things worth flagging. First, AI safety researchers are openly nervous. A two-to-three week release cycle doesn't leave much room for red-teaming a 2 trillion parameter model. Second, CIOs are watching but not committing — they want stability and API reliability before they build on this. Speed-to-market is a real edge, but it's also a real liability if something goes wrong in production. And the connection back to Bitcoin is more direct than it looks — Hayes and others are arguing that this compute buildout is being financed on credit that eventually has to be inflated away.
Now to the story that's rattling Bitcoin self-custody. The Coldcard exploit has grown into a roughly $130 million incident. Galaxy Research confirmed three major attack waves plus 14 smaller incidents, and they're tracking a possible fourth wave that moved another 448 BTC from about 709 addresses. At least 15 distinct attackers have exploited the flaw — which, per Dragonfly's managing partner, could have been prevented with about $2 worth of hardware entropy hardening.
The root cause is a five-year-old firmware bug in how seeds were generated. Coinkite says new seeds on patched firmware are safe, but anything created on vulnerable firmware needs a full migration. That's a big ask for users who bought a hardware wallet specifically to not have to think about this again.
And here's the awkward part. Cory Klippsten and others are framing this as an accelerant for collaborative multisig. But the more immediate market reaction is that some Bitcoin is flowing back toward regulated custodians and ETFs. Spot Bitcoin ETFs pulled in $382 million over two days, with seven different funds going positive simultaneously after a brutal $265 million outflow day. Galaxy's own ETF returned to gains.
The hacker's wallet, holding around $36 million in stolen coins, has turned into a bizarre public message board — victims paying transaction fees just to leave permanent pleas on-chain asking for some of it back. It's grim, but it's also a reminder of what final settlement actually means.
The honest read: hardware wallets are still the right default for most self-custodians, but single-signer setups are looking weaker every year. If you're holding meaningful size, this is the nudge to look at 2-of-3 multisig across different vendors.
Corporate Bitcoin treasury strategies are splitting into two very different camps, and this week made that obvious.
Metaplanet added 2,823 BTC in Q2 at an average of roughly $78,850 per coin, pushing total holdings past 43,000 BTC — about $4.1 billion. Their blended cost basis actually came down to $95,117 as they bought weakness. And crucially, they generated around $10.95 million in quarterly revenue from selling cash-secured options against the stack. That's an active yield strategy, not just a static reserve.
Bhutan is going further. Their Gelephu Mindfulness City is handing part of its 10,000 BTC stash to 3iQ, a regulated Canadian manager, to run a market-neutral strategy. A sovereign entity paying an asset manager to generate yield on Bitcoin — that's a real signal about how mature institutions now view this asset class.
Matador got Canadian regulatory approval for a $58 million share offering specifically to fund Bitcoin accumulation. Hyperscale Data went a different direction — they're borrowing about $30 million against their Bitcoin at 4.9% through Morpho to fund an AI data center in Michigan. Bitcoin as collateral for AI infrastructure. That's the loop closing.
But the other camp is monetizing. K Wave Media sold its last 88 BTC to pay down debt. Sequans said it'll drawdown its holdings over time. And Strategy — yes, that Strategy — filed disclosure of selling 1,638 BTC between July 27 and August 2 for $104.7 million, splitting proceeds between dividends and share repurchases. They still hold about 842,000 coins, but the pure accumulation narrative just cracked a little.
GameStop is planning a $1.4 billion note swap that would cut principal by a third, with Bitcoin option terms undisclosed — a reminder that some of these treasury plays were financed with structures that now have to be renegotiated.
The pattern: firms with real cash flow and conviction are layering yield strategies on top of accumulation. Firms that levered up in the enthusiasm phase are now being forced to sell or restructure. Cycle behavior, playing out in slow motion on public balance sheets.
Now the market picture. Bitcoin is sitting around $63,800, right on top of its 200-week moving average at $63,657. Glassnode data shows about 515,000 BTC — roughly 3% of circulating supply — is concentrated in the $63,000 zone. That's the second densest cost-basis cluster after the $78,000 to $82,000 range. This is a genuine battleground level.
On-chain, the accumulation signal is loud. About 155,000 BTC moved into the $62,000 to $65,000 cost basis during the recent pullback. Wallets holding 10 to 10,000 BTC — call them whales and mid-tier holders — added 19,610 coins. Retail wallets under 0.01 BTC shrank by 0.55%. The Adaptive Sell-side Risk Ratio dropped to the third percentile, meaning very little profit-taking is happening.
That's the constructive read. The cautious read: spot volumes are the lowest since late 2023. Implied volatility is near multi-year lows. And Glassnode says its aggregate price-cycle tool is in its coldest phase since the FTX collapse. Translation — nobody is excited, nobody is panicking. Just grinding sideways between $60,000 and $67,000.
Which brings us to Arthur Hayes. His new thesis is that the AI infrastructure buildout — the same 15 gigawatts of compute Musk was talking about yesterday — is being financed with debt that will eventually strain. When it cracks, governments bail out the hyperscalers, print money, and Bitcoin runs to $1 million. It's the same crack-up boom argument he's been making for years, now dressed in AI capex. Worth noting: evidence of actual financial strain across Big Tech is uneven. Hyperscalers still have monster free cash flow. But the leveraged AI-adjacent players — the miners pivoting to compute, the smaller data center operators — those are the ones to watch.
And one macro data point that matters. Texas Governor Abbott just froze 474 gigawatts of pending data center grid connection requests pending audit. That's an absurd number — larger than the entire Texas grid. It suggests the AI power-demand story is running into physical and political limits faster than the models suggest.
One prediction: within 18 months, single-signer hardware wallets will feel as dated as password-only email accounts. The Coldcard incident didn't kill self-custody — it just raised the floor on what serious self-custody looks like.