Open Source AI — October 2, 2026
Yesterday's Nebius price increases (effective today) crystallize a diverging market: model prices are plummeting while the cloud infrastructure needed to run them is getting more expensive.
In Plain English
Yesterday's Nebius price increases (effective today) crystallize a diverging market: model prices are plummeting while the cloud infrastructure needed to run them is getting more expensive. GPT-6 Sol costs 50% less than its predecessor, yet GPU instances and memory jump 17-41% in price - a widening gap that accelerates the shift toward on-prem and local ownership models running open-weight LLMs.
Nebius Raises GPU Rental Rates 17-41% Effective October 1; Memory Surge Signals Hardware Scarcity
Cloud GPU provider Nebius implemented significant price increases today following an announcement on September 23. Nvidia H100s and H200s rose 17-20%, B-series GPUs climbed 19-21%, AMD EPYC instances jumped 25%, and memory-backed instances surged 41% - the largest increase in the price hike. Nebius attributed the increases to "continued strong demand for AI computing power outstripping available supply," reflecting upstream semiconductor constraints on high-bandwidth memory and DRAM. The 41% memory jump is particularly telling: as autonomous agents and agentic systems proliferate, they consume vast amounts of working memory, straining hardware availability precisely where it hurts most.
Infrastructure & Economics · Shattered.io · September 23, 2026 (Effective October 1)
Price Divergence Widens: Model Deflation Meets Infrastructure Cost Inflation; Margin Pressure Accelerates
The 404K Semi-AI brief documents a stark market split emerging in late September: foundation model APIs are entering a race to the bottom while cloud infrastructure costs accelerate upward. OpenAI's GPT-6 Sol slashed pricing 50% to $2/$10 per million tokens, with lightweight Luna at $0.10/$0.50; Anthropic cut Claude Opus 5.5 by 20% to $4/$20 per million tokens. Meanwhile, Nebius GPU hikes, memory fab constraints, and autonomous agent proliferation are driving infrastructure costs higher. The brief warns that application developers now face a brutal squeeze: they buy compute at falling model prices but rent it from clouds at rising infrastructure rates, compressing margins on both ends.
Infrastructure & Economics · 404K Research · September 23, 2026
The Apple Angle (Standing Perspective)
Yesterday's Nebius increase and the widening gap between falling model prices and rising infrastructure costs reframe one quiet advantage. A Mac Studio with 512GB unified memory (~$9,500) runs any open-weight model released in the last six weeks - entire trillion-parameter models - at zero per-token cost. No Nebius invoice. No GPU rental. No margin squeeze between deflating model prices and inflating cloud compute. Five to six Nvidia RTX Pro cards to match that memory footprint cost $60,000-$75,000 and consume ten times the power. As open-weight releases from Qwen, DeepSeek, Mistral and others hit production-ready capability every few weeks, the economic case keeps tilting: why rent compute from Nebius at rising rates when Apple's unified memory approach lets you own the box outright? Apple's MLX framework and Thunderbolt 5 RDMA for multi-machine pooling are the quiet infrastructure that makes local ownership look increasingly rational - not as a downmarket alternative, but as the fundamentally more cost-efficient answer to the margin-squeeze problem that hits everyone else. (Caveat: Mac Studios serve single-user local inference well; megawatt datacenters still serve millions. But on the margin where the tradeoff is between "rent my compute from a rising-cost cloud provider" vs. "own a Mac Studio," Apple's story looks cleaner every day.)
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
The Nebius price increase is not an isolated event - it's a visible marker of the underlying squeeze that Lucien's thesis predicts. Model prices are in free fall (down 50-90% over 12 months for frontier models), driven by competition and open-weight alternatives. But the infrastructure that runs them - memory, GPUs, power, cooling - is tightening and getting more expensive. The margin gets crushed in the middle. Yesterday's brief documented that Anthropic is burning $42 billion per year while raising another $518 billion for capex, betting that consumption will scale past profitability. Today's Nebius hike shows the opposite pressure: if you're an application developer buying cheaper models but renting infrastructure at rising rates, your ROI timeline extends or your margin disappears. This is exactly when the decision to shift from "rent everything from the cloud" to "own a box and run open-weight models locally" becomes not exotic, but rationally mandatory. Every Nebius rate hike, every token price cut, pushes that calculation further toward local ownership. Watch the next: as more enterprises run breakeven analyses like the Lenovo whitepaper mentioned yesterday, adoption of on-prem and on-device inference will accelerate, not as a downgrade, but as a response to the simple math that ownership is cheaper.
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.