Open Source AI — September 26, 2026
Three stories today, all showing how expensive and complicated it's getting to run AI the conventional way. Electricity, not chips, is now the bottleneck slowing new AI data centers, and companies are scrambling to buy backup generators just to keep the lights on.
In Plain English: Three stories today, all showing how expensive and complicated it's getting to run AI the conventional way. Electricity, not chips, is now the bottleneck slowing new AI data centers, and companies are scrambling to buy backup generators just to keep the lights on. Anthropic signed a deal worth up to $20 billion for computing capacity, handing over a slice of company stock as part of the arrangement - a sign of how much cash this all still takes. Meanwhile, new research shows companies are already routing over a third of their AI work through free, downloadable models instead of paid ones, and that share is climbing fast because it's cheaper. Today's Apple Angle explains why owning the hardware sidesteps all three problems at once.
Electricity, Not Chips, Is Now the Real Constraint on the AI Data Center Boom
New estimates say AI chip electricity demand could jump more than 1,100% from 2025 levels to roughly 315 gigawatts by 2033, with the US alone needing about 200 of those additional gigawatts - even as hyperscaler capex is projected at roughly $916 billion over the next 12 months and nearly $1.2 trillion the year after. Meta, Amazon, Google and Oracle are now signing multibillion-dollar backup-generator deals and grid-upgrade pledges just to keep pace, since AI workloads can spike power draw as much as 50% above a facility's design capacity. For Lucien's thesis, this is prediction #1's cost pressure taking a new, physical form: even if credit markets stayed calm, the buildout now runs into a hard grid-capacity ceiling that money alone can't solve quickly.
Infrastructure & Economics · Benzinga · Sep 25, 2026
Anthropic Commits Up to $20 Billion to Akamai for CPU Capacity - and Akamai Hands Over a Stake in Return
Anthropic signed a seven-year, $11.6 billion cloud deal with Akamai that could grow to roughly $20 billion, mainly for CPU capacity to handle the code-execution and browsing work its AI agents increasingly generate - Akamai's largest contract ever, expected to reach a $1.7 billion annual run rate by 2028. In an unusual reversal of the industry's typical pattern, Akamai is granting Anthropic a warrant convertible into up to 5% of its stock, vesting in tranches tied to how much Anthropic actually spends. For Lucien's thesis, this is prediction #1's capex escalation continuing unabated at the model-lab level, with an equity-for-commitment structure that looks less like an ordinary customer contract and more like financial engineering to lock in years of demand.
Infrastructure & Economics · TechCrunch · Sep 25, 2026
New Research: Open-Weight Models Are on Track to Handle 41% of Enterprise AI Tokens
Enterprise Technology Research finds open-weight models already account for 34% of enterprise AI token usage, up from 23% a year ago, with adoption projected to reach 41% as more companies move workloads into production - deployment rates jumped to 42% in September from just 31% in July. Among companies already running open-weight models in production, the share processing more than half their tokens that way jumped from 3% to 24% in two months, and 69% of production users cite cost as their primary motivation, with many willing to switch for savings as small as 25%. For Lucien's thesis, this is prediction #3 showing up as measured enterprise behavior rather than anecdote - real token volume moving off commercial APIs, driven by exactly the price sensitivity prediction #1 anticipates.
On-Prem/On-Device Shift · Techstrong.ai (Enterprise Technology Research) · Sep 22, 2026
The Apple Angle (Standing Perspective)
Today's news is really three variations on the same complaint: centralized AI compute is running short on power, requires ever-larger and stranger financing structures to secure, and costs enough that enterprises are already voting with their token volume. Apple's answer sidesteps the whole chain. A Mac Studio configured with its full 512GB of unified memory - about $9,500 - holds an entire trillion-parameter open-weight model in memory on a single desktop machine, drawing a fraction of a rack's worth of power and needing no grid-interconnection study, no warrant negotiation, no seven-year contract.
Reaching that same memory footprint on Nvidia's own workstation cards takes five or six RTX Pro 6000s, somewhere around $60,000 to $75,000 combined, while pulling roughly ten times the electricity - the very resource today's lead story says is now the limiting factor. As frontier-capable open-weight models keep shipping every few weeks, the interesting competitive question keeps drifting away from who trained the best model and toward who can simply afford, and power, the hardware that runs a nearly-as-good one at home. Apple's MLX software and Thunderbolt 5's RDMA networking already let multiple Mac Studios pool memory into one larger effective machine, Nvidia has dropped NVLink from its own consumer workstation cards, and its DGX Spark leans on a unified-memory design closer to Apple's than to Nvidia's traditional GPU architecture. The honest caveat still applies: a Mac Studio serves one person running one model at a time, while the power-hungry, multibillion-dollar facilities in today's stories exist to serve millions of concurrent users at once - Apple looks positioned to quietly own a meaningful slice of AI compute, not to replace the data center outright.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's items apply pressure from three directions at once - a literal power ceiling limiting new data center capacity, a major AI lab locking in years of spending through an unconventional equity-for-commitment deal, and enterprises already shifting real token volume onto free, open-weight models to escape those very costs. That's prediction #1's cost squeeze and prediction #3's flight toward ownership showing up in the same week, exactly the kind of shift the standing Apple Angle keeps pointing toward.
The Daily: Open Source AI - a recurring research brief for Lucien Engelen
Book Lucien for your next event
Keynotes, masterclasses, panels and board-room sessions.