The Hermes Dispatch | August 25, 2026
4 min read | TL;DR: Anthropic gives Claude a real memory, Stability AI raises another $76M, and Bitcoin miners are becoming AI landlords.
The Rig
Agent TL;DR: NVIDIA's RTX 5090 now sits at the top of the 2026 local-LLM hardware stack with 32GB of GDDR7 VRAM, but Mac Studio M4 Max unified memory still wins on bandwidth and simplicity.
Kunal Ganglani's latest 2026 hardware guide ranks the RTX 5090 as the largest consumer VRAM option at roughly $1,999 MSRP, making it the go-to for running 32B-parameter models locally without cloud tokens. It is followed by the RTX 4090 at 24GB and the RTX 4060 Ti 16GB for mid-range 14B workloads. For builders on a budget, an RTX 4060 or Intel Arc B580 at around $300 can comfortably run 7-8B models, and at 1,000+ API calls per month the economics already beat cloud APIs.
The bigger architectural story is unified memory. No consumer NVIDIA card can match the 128-192GB of unified memory available on Apple Silicon, which matters because memory bandwidth, not just raw compute, is often the bottleneck for inference. YouTuber Kai made the same point in a recent hardware breakdown: the right inference framework can matter as much as the GPU itself. The DGX Station GB300 Blackwell is now supported by both Ollama and LM Studio, while llama.cpp and vLLM continue to split the ecosystem between ease-of-use and throughput.
Why it matters: Local inference is moving from hobbyist project to production backend. By early 2027, an 8B model on a $300 GPU is expected to match what GPT-4 could do in 2024, which turns the old "cloud vs local" cost debate upside down for privacy-sensitive and high-volume users.
The play: If you are building a personal or small-team AI workstation, start with the RTX 5090 or Mac Studio M4 Max depending on whether your workflow values raw VRAM or unified bandwidth. Match it to the right stack: Ollama or LM Studio for experimentation, vLLM for serving, and llama.cpp for broad model compatibility.
The Mine
Agent TL;DR: Public Bitcoin miners are pivoting from hash-rate sales to long-term AI data-center leases, with CoinShares projecting AI revenue will exceed 80% of total revenue at contracted miners by year-end.
Bitcoin mining stocks have outperformed in 2026 even as Bitcoin itself struggled, because investors are repricing them as AI infrastructure plays. Galaxy Research estimates U.S. data center demand will hit 45 GW by 2030, up from 21 GW in 2024, and legacy data centers cannot retrofit fast enough for 132 kW racks like the GB200 NVL72. Listed miners already control more than 27 gigawatts of planned power capacity, and announced AI deals cover about 3.7 GW of that, roughly 14%.
Companies furthest along include IREN, TeraWulf, Cipher, Core Scientific, Hut 8, and Applied Digital. IREN is now effectively an AI cloud provider with a multi-billion-dollar Microsoft contract and a direct Nvidia partnership, projected to get roughly 71% of revenue from AI and high-performance computing by the end of 2026. Bernstein estimates miners still trade at about a 90% discount to established data center operators on certain metrics, despite the recent rally.
Why it matters: This is the most practical crypto-meets-AI convergence in public markets. AI contracts generate roughly three times the revenue per megawatt as traditional mining, with predictable cash flows and credit-worthy hyperscaler tenants.
The play: If you hold mining exposure, look beyond Bitcoin price to contracted AI megawatts and the quality of the offtake partner. If you are new to the sector, the CoinShares Bitcoin Mining ETF (WGMI) offers broad exposure to the pivot.
Secure mining payouts with Ledger →
The Ledger
Agent TL;DR: Schwab and Robinhood are both racing to put AI inside the investing workflow, with Schwab launching AI-powered portfolio explanations and Robinhood rolling out agentic trading accounts.
Charles Schwab rolled out its first generative AI capability for retail clients in May 2026, starting with AI-powered summaries of portfolio performance for up to five holdings that moved the most since the prior day. Schwab surveyed nearly 1,000 retail clients and found over 60% are interested in using AI, while nearly 70% believe it can play a meaningful role in investing when paired with human expertise. The firm is also expanding into concentration risk, asset allocation, and technical indicators.
Robinhood is taking a different path: "agentic trading" gives an AI agent a dedicated account to trade in, with users monitoring activity and performance directly in the app. Corporate Insight flagged this as a broader trend in an August 2026 report titled "Fintechs Put AI in the Driver's Seat with Agentic Trading," noting that fintechs are moving past chatbot summaries and letting AI actually execute orders.
Why it matters: Retail brokerage is becoming a battleground between human-in-the-loop AI explanations and autonomous-agent execution. Schwab is betting clients want context first; Robinhood is betting some will let the machine trade.
The play: Start with AI summaries and performance explanations to understand what the tools get right and wrong before you hand over execution authority. Use a broker that gives you both education and control, not hype.
Quick Bites
- Anthropic is giving Claude a shared memory across chat and Cowork, so users no longer have to repeatedly brief the AI on projects and preferences.
- Stability AI, the maker of Stable Diffusion, raised $76 million in fresh funding, bringing its total fundraising to $232 million.
- Gamma acquired Accel-backed design startup Lica, and Lica's co-founders are joining Gamma's new research team.
- Self-driving truck startup Gatik raised $200 million, led by Qatar Investment Authority and Koch Disruptive Technologies, following a PepsiCo deal.
⚙️ Mission Freedom: Behind the Scenes
- What we shipped: The HermesDispatch site at hermesdispatch.dev stayed reachable with its subscribe CTA, lead magnet, and tools index all live. Last night's newsletter,
MF-20260824-001, was generated and sent successfully to 1/1 subscribers via Resend, and the site was updated with the new issue. The Overnight Learning Orchestrator analyzed 100 runs across 44 domains with a 0.0% failure rate, and the overnight Windows migration completed cleanly. SENTINEL's weekly audit logged 0 critical, 0 high, and 0 warning findings out of 13 total checks. - Current experiment: Tuning the newsletter send window so
newsletter_orchestrator.pydoes not skip the 10:00 UTC attempt again; the cron is set to 16:04 MDT but the orchestrator also checks for the UTC hour 22 ± 30 minute window. - What's next: Resolve the DNS issue for
hermesmissionfreedom.aiwith the registrar/DNS setup, then build out the hardware compare page referenced in today's Rig CTA so readers can filter GPUs by VRAM, price, and framework compatibility.
The Hermes Dispatch is written by dare404 in Boise, ID and generated with AI assistance.
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.
Sources: Reuters AI News, AI Weekly, TechCrunch, Kunal Ganglani local-LLM hardware guide, Galaxy Research, CoinShares, Corporate Insight, Charles Schwab press release, Robinhood.
Generated: 2026-08-25 16:01 MDT