Open Source AI — August 22, 2026
Two stories today, and together they sketch both halves of the thesis.
In Plain English: Two stories today, and together they sketch both halves of the thesis. Broadcom is trying to raise up to $100 billion in new debt — an enormous, leveraged bet — just to build the AI computing capacity Anthropic needs, which is exactly the kind of spending investors worry won't pay for itself. Meanwhile, Google announced its free, open-weight Gemma models have now been downloaded over a billion times and turned into more than 100,000 variants, including one running on a NASA satellite — proof that "good enough" open models are already spreading into places commercial APIs can't easily reach. The standing Apple argument below explains why that spread matters even more as compute financing gets this stretched.
Broadcom Seeks Up to $100 Billion in Debt to Fund Anthropic's AI Computing Buildout
Broadcom is negotiating a debt package that could reach $100 billion — split into a $60–70 billion senior tranche and a roughly $30 billion junior tranche, routed through a special-purpose vehicle and backed in part by lenders Apollo Global Management and Blackstone — to finance AI chip capacity for Anthropic. The deal builds on a June agreement targeting more than 20 gigawatts of AI compute by 2028, and analysts note it stacks a lot of leverage on the bet that hyperscaler AI spending keeps climbing; if that spending slows, the debt-heavy structure becomes a lot more exposed.
Yahoo Finance (Bloomberg) · Infrastructure & Economics · Aug 20, 2026
Google's Free, Open-Weight Gemma Models Pass 1 Billion Downloads and 100,000 Variants
Google DeepMind says its open-weight Gemma model family has now been downloaded more than a billion times, with developers building over 100,000 distinct fine-tuned variants. Real-world deployments cited include NASA's Jet Propulsion Laboratory running a compressed 4-bit Gemma model directly on a satellite for image analysis (88% accuracy), and India's National Health Authority embedding Gemma in a health app with 100+ million downloads — concrete evidence that open-weight models are already running on edge devices and inside government systems, not just in research demos, without anyone paying a commercial API bill.
Unite.AI · Open-Weight Models · Aug 20, 2026
The Apple Angle (Standing Perspective)
Worth repeating alongside today's numbers: while Broadcom leverages itself to the hilt building data-center capacity, a different, quieter hardware bet has been sitting in plain sight — Apple's. Because Apple Silicon pools memory across the whole chip instead of splitting it per card, a single Mac Studio, priced around $9,500, can be configured with up to 512GB of unified memory — enough to run a trillion-parameter open-weight model on one desktop machine. Nvidia's RTX Pro 6000 workstation cards would need five or six of them, roughly $60,000–$75,000 total, to match that capacity, while drawing about ten times the power.
The argument gets stronger with each open-weight release, today's Gemma milestone included: as capable open models keep shipping every few weeks from Google, DeepSeek, GLM, Kimi, and others, the constraint shifts away from "which lab has the best model" and toward "who can afford to run it locally." Multi-Mac clusters using MLX and Thunderbolt 5's RDMA networking can already pool memory across several Studios for even larger models, and Nvidia's own DGX Spark essentially validates the unified-memory approach — telling, since Nvidia has stripped NVLink pooling out of its consumer and workstation cards. The fair caveat remains: a Mac Studio serves one user running one model locally, while a data-center GPU cluster serves many people simultaneously — so this is Apple leading a specific, valuable segment, not the entire AI compute market.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's two items sit on opposite sides of the same coin. Broadcom's up-to-$100-billion debt package to build capacity for Anthropic is precisely the kind of leveraged, "profitability assumes spending keeps climbing" bet that raises the question of whether current AI infrastructure investment can ever pay for itself as inference economics tighten. At the same time, Gemma crossing a billion downloads and 100,000 variants — running everywhere from a NASA satellite to a 100-million-user health app — shows open-weight models are no longer a research curiosity but an already-deployed alternative that costs nothing per token to license. Set against the standing Apple argument, the throughline holds: the more debt-financed and leverage-dependent commercial AI infrastructure becomes, the more attractive a one-time hardware purchase running a free, capable open-weight model locally looks by comparison.
Book Lucien for your next event
Keynotes, masterclasses, panels and board-room sessions.