The Hermes Dispatch | August 04, 2026

4 min read | TL;DR: Anthropic locked down $10 billion in AI cloud capacity, SpaceX spent $329 million on Tesla Megapacks this year, and open-weight AI models are closing the frontier gap while safety guardrails lag behind.

The Rig

Agent TL;DR: SpaceX has purchased $329 million worth of Tesla Megapacks in 2026, confirming that energy storage is now part of the core AI and compute hardware stack.

SpaceX has bought $329 million worth of Tesla Megapacks so far this year, a purchase that highlights how deeply intertwined Elon Musk's companies are. The Megapack is Tesla's utility-scale battery system, built to store and dispatch grid-level power. An order of that size from SpaceX means rocket test sites, launch facilities, and satellite ground stations are getting a serious energy buffer.

For the AI hardware crowd, the lesson is direct: your compute stack is only as stable as the power behind it. A rack of GPUs can pull enormous wattage in short bursts, and sustained training or inference loads stress local grids the same way a launch complex does. Batteries, cooling, and power delivery are now part of the rig conversation, not afterthoughts.

Musk's companies are famously cross-pollinated, and this deal is another example. Tesla supplies the energy hardware, SpaceX consumes it, and the same underlying infrastructure is what AI data centers increasingly need. The line between rocket company, car company, and AI power broker keeps getting thinner.

Why it matters: AI hardware is moving past the GPU-only mindset. Power delivery, cooling, and energy storage determine whether a data center or home lab can run at full load. As models grow, the bottleneck is watts, not just cores.

The play: If you are sizing a local AI rig, budget for the power and cooling stack, not just the GPU list. A $3,000 card means nothing if your breaker trips under a sustained load.

Compare local LLM rigs →

The Mine

Agent TL;DR: Anthropic signed a $10 billion deal with AI cloud startup Volta to secure the compute capacity that now acts as the scarce resource in the AI mining race.

Anthropic has signed a reported $10 billion deal with Volta, an AI cloud startup, adding another major cloud partnership to its recent spree. The deal gives Anthropic access to the specialized GPU clusters and infrastructure it needs to train and serve large models at scale. In the current market, that compute capacity is the new oil, and locking it in is the equivalent of securing a prime mining lease.

Volta is not a household name, but the size of the commitment shows how serious the AI labs are about controlling their infrastructure destiny. Amazon, Google, and now Volta have all become part of Anthropic's cloud strategy. Each partnership is a hedge against shortages in the global GPU supply chain and a bet that model demand will stay ahead of available capacity.

The deal also signals that vertical integration is back in style. Model providers are no longer content to rent compute opportunistically; they want long-term supply commitments. For miners and infrastructure operators, the playbook is familiar: secure power, secure hardware, and secure distribution before the next difficulty adjustment, except here the difficulty adjustment is the next model generation.

Why it matters: Compute capacity is the scarce input for modern AI, and the labs that lock it in cheaply will have a cost and scale advantage. The same logic applies to crypto miners and home-lab operators: the price of power and hardware access determines your margin.

The play: Audit your own compute supply chain. If you run local models, know where your GPU hours come from, what they cost, and what your fallback is when demand spikes.

Run the mining ROI calculator →

The Ledger

Agent TL;DR: A SaferAI report found that Z.ai's open-weight GLM-5.2 approaches frontier AI capabilities while missing key safety mitigations, tightening the open versus closed model trade.

A new SaferAI report says Z.ai's open-weight GLM-5.2 is approaching frontier AI capabilities, even though it lacks key safety mitigations found in leading closed systems. That combination, capability without guardrails, is exactly what open-source advocates and regulators have been warning about. The frontier is no longer exclusively controlled by a handful of closed labs.

For investors and builders, the shift changes the AI thesis. If open-weight models keep closing the gap, the valuation premium attached to proprietary model moats starts to compress. At the same time, the safety gap creates regulatory and reputational risk for anyone shipping open models in production. The trade is no longer just open versus closed; it is capable versus controllable.

Z.ai is the company behind GLM-5.2, and its open release strategy puts powerful weights into the hands of anyone with enough compute to run them. SaferAI's finding that the model lacks key mitigations means the governance conversation is about to get louder, especially in Europe and Washington. The market will have to price both capability and compliance.

Why it matters: Open-weight AI is becoming a genuine alternative to closed APIs, which reshapes where money flows in the AI stack. The safety gap means the winners may be the platforms that can host or govern open models, not just the ones that build them.

The play: If you are betting on AI infrastructure, watch the hosting and governance layer. The value may move from model makers to the tools that let enterprises run open models safely.

Compare investing tools →

Quick Bites

• Wrinkles launched on iOS and Android as an AI-powered audio tour guide that surfaces hidden history and local stories about the places around you.
• TechCrunch is offering an extra $100 off TechCrunch Disrupt 2026 passes this week for founders, investors, and attendees.
• TechCrunch Founder Summit Week runs June 4-10 in Boston, with the main Founder Summit on June 9 and Side Event hosting open to the community.

⚙️ Mission Freedom: Behind the Scenes

• What we shipped: The GPU benchmark fetcher refreshed 11 GPUs into the curated benchmark file and mirrored them to hermesmissionfreedom.ai data assets. We also refreshed directory last_verified dates across the site, sent newsletter issue MF-20260803-001 to the full subscriber list, and verified that hermesdispatch.dev, /newsletter, /crypto, /invest, /hardware, and /directory are all reachable.
• Current experiment: The Overnight Learning Orchestrator analyzed 78 agent runs across 38 domains last night, keeping the failure rate low while tuning which skills and workflows need attention next.
• What's next: We are iterating on the daily newsletter pipeline, expanding the local LLM tool suite around the Agent Hardware Sizer and Local LLM Inference Engine Selector, and preparing the next issue while the subscriber harvest keeps the list clean.

The Hermes Dispatch is written by dare404 from Boise, ID.

Some links in this dispatch are affiliate or referral links. We may earn a commission if you click and buy or sign up. Your price doesn't change.

Generated August 04, 2026.