Bitcoin's flirting with 65,000 again after the softest U.S. inflation print since 2020, and rate-hike odds just collapsed from 43% to 13%. OpenAI dropped a three-tier GPT-5.6 lineup on Amazon Bedrock, branded Sol, Terra, and Luna. MARA is spending up to 600 million dollars to turn a Texas e-fuels site into a 2 gigawatt AI and mining campus. And Stripe, teamed up with Advent, just lobbed a 53 billion dollar bid at PayPal. Let's get into it.
OpenAI just made its strongest model family generally available on Amazon Bedrock, and they've ditched version numbers for names: Sol, Terra, Luna. Same GPT-5.6 generation underneath, three very different price and performance points.
Sol is the flagship reasoning model. The benchmark numbers are the story here. On the Artificial Analysis Coding Agent Index, Sol hits 80. On ExploitBench, which measures cybersecurity capability, it jumps to 73.5% versus 47.9% for GPT-5.5. And it's doing this at roughly one-third the cost of the prior flagship, with lower token usage and faster response times. If those numbers hold up in the wild, that's a serious jump in autonomous coding, vulnerability research, and multi-step agent work.
Terra is the middle tier. Better than GPT-5.5, cheaper than GPT-5.5, aimed at the boring but massive workloads — code generation, structured data extraction, content pipelines, general agent tasks. This is where most enterprise spend actually lives.
Luna is the small, fast, cheap one. Classification, summarization, routing, real-time apps. High volume, low latency.
The Bedrock side of this matters too. AWS is pushing prompt caching with a 90% discount on cached input tokens, retained for at least 30 minutes. For multi-step agent workflows that reuse the same context over and over, that's a massive cost cut. Add in-region inference, zero-operator access, IAM enforcement, VPC deployment, CloudTrail logging — this is OpenAI's models wrapped in AWS enterprise controls.
Sol is only in US East for now. Terra and Luna are in US East and US West. The strategic read: OpenAI is moving to capability-tiered branding and leaning on AWS to reach enterprises that were never going to hit OpenAI's API directly. And on the same day, Google's Gemma-4-E2B-it landed on SageMaker JumpStart — multimodal, function-calling, instruction-tuned. AWS is aggressively becoming the neutral ground where every frontier model sits behind one billing relationship.
While the frontier labs are shipping bigger models, the security industry is quietly shipping something more interesting: autonomous AI agents that actually take action inside your infrastructure.
Blackpoint Cyber just made its AI SOC Agent generally available for identity threat detection. The pitch: it detects and contains credential-based attacks on Microsoft 365 and Google Workspace in an average of under 2 minutes, with fastest cases at 21 seconds. Token theft, impossible travel, business email compromise — the agent investigates, decides, and contains, with a human analyst providing oversight on the escalations. It was trained on years of SOC decisions and telemetry from nearly a million accounts. Microsoft reported a 32% jump in identity-based attacks in the first half of 2025, so the timing is not accidental.
Smarttech247 launched something similar inside its VisionX platform — agentic SOC built specifically for the Microsoft Defender and Sentinel stack. They cite identity-related incidents as roughly 80% of their investigated cases this year.
And here's the one that flips the script: Codenotary released AgentMon 3, which is security for AI agents themselves. Their bet is that as companies deploy autonomous coding assistants and internal AI systems, those agents become the new attack surface. AgentMon watches actual runtime behavior — file access, network calls, credential use — not what the agent claims it's doing. It analyzes over 5 million agent interactions daily, cryptographically logs every decision, and auto-tunes policies. They claim 80% less manual policy tuning.
Put these together and a pattern emerges. AI defending against AI-generated attacks. And AI watching other AI. This is the actual agentic workflow story — not chatbots writing emails, but autonomous systems making containment decisions at machine speed while humans handle the judgment calls. The MSPs and MDR providers who don't ship something like this in the next 12 months are going to look very slow.
MARA Holdings is buying a 1,200-acre powered land site in Matagorda County, Texas, about 90 miles southwest of Houston, for up to 600 million dollars in milestone-based payments. The seller is HIF USA, which had originally planned the site as a synthetic e-fuels facility. Now it's becoming a digital infrastructure campus for high-performance computing and Bitcoin mining.
The scale is what matters. Up to 1 gigawatt of grid capacity by October 2027. Up to 2 gigawatts by April 2028. Combined with MARA's pending Long Ridge acquisition in Ohio, this pushes the company's total power portfolio toward roughly 4.8 gigawatts. That's utility-scale.
And notice the framing. MARA isn't pitching this as a mining expansion. They're partnering with Starwood Digital Ventures, who handles design, construction, and tenant sourcing. HIF keeps a minority stake, but only once a high-performance computing lease is signed. The tenant mix is deliberately undecided — AI workloads, flexible compute, mining, whichever pays best. This is a Bitcoin miner explicitly repositioning as a digital energy landlord.
But there's a catch, and it's showing up on the same week. ERCOT — the Texas grid operator — just released new large-load interconnection rules aimed squarely at industrial power users like miners and hyperscalers. The mining economics in Texas are no longer just about cheap power and demand response payments. They're now tied directly to how ERCOT decides to gate new interconnections. Miners who thought they could just plug in gigawatts on demand are going to find the rules meaningfully tighter.
Meanwhile, CleanSpark's stock jumped 22% on a 6.6 billion dollar Georgia data center lease with an undisclosed investment-grade tech tenant. Bitmine, which pivoted from Bitcoin mining to Ethereum staking, generated 46 million dollars last quarter — 98% of revenue — from ETH staking after launching validators in March.
The miners are splitting into two camps. One camp — MARA, CleanSpark, Riot — is becoming a power-and-real-estate business selling capacity to whoever wants it, AI included. The other camp is quietly exiting Bitcoin mining entirely. The pure-play publicly traded Bitcoin miner as a category is being hollowed out from both sides.
June CPI came in at the softest reading since 2020. Rate-hike odds for the next Fed meeting collapsed from 43% to 13% within hours. Bitcoin pushed to 64,800, within 200 dollars of 65,000. Some technical analysts are now floating an 80,000 target for August, though there's a matching bear-case camp warning about a 2022-style rerun if inflation reaccelerates.
The ETF flows told the story. U.S. spot Bitcoin ETFs pulled in 181 million dollars on Tuesday. Ether ETFs added 58 million. Combined July 14 inflows across both categories hit 239 million — reversing roughly two months of outflows totaling over 8 billion dollars since May. BlackRock's IBIT and ETHA were the big beneficiaries.
But zoom out and the picture is more nuanced. BlackRock's total digital asset holdings actually fell 39% over the past year despite 15 billion in net inflows, because prices dropped harder than money came in. ETF trading volumes are still running about 78% below their peak per Glassnode. And on Monday — the day before the CPI print — Bitcoin ETFs saw 425 million dollars in outflows led by Fidelity and BlackRock. This isn't a clean trend yet. It's a market that's twitchy, waiting for the September FOMC.
The more interesting data point: a 1.3 billion dollar block of IBIT changed hands in a dark pool trade. Off-exchange, opaque, no counterparties disclosed. That kind of liquidity event tells you institutional players are moving size around outside the visible order books.
And on the geopolitics side — the U.S. Treasury sanctioned four Tron wallets holding roughly 131 million dollars in USDT tied to Iran. Total Iran-linked freezes via Tether now sit near 475 million dollars over three months. Whatever your view of stablecoins as neutral infrastructure, that ship has sailed. USDT is now an active tool of U.S. sanctions enforcement, and Tether is cooperating. Middle East tensions are the reason Bitcoin didn't run harder on the CPI print — every bullish macro headline is getting partially offset by oil and Iran risk.
One more to watch: Mizuho downgraded Circle to underperform with a 50 dollar price target, citing the coming Open USD consortium stablecoin that would share reserve income with distributors rather than keeping it at the issuer. If Open USD ships in 2026 as planned, Circle's margin structure gets squeezed hard. Cathie Wood's ARK apparently disagrees — they bought another 220,000 Circle shares this week.
Here's the pattern worth watching. The frontier model tiers, the autonomous SOC agents, the miners becoming power landlords, the ETF flows moving through dark pools — every one of these stories is about infrastructure hardening into tiers and choke points. The retail-friendly narrative is thinning out. The plumbing is where the next few years actually get decided.