Open Source AI — October 10, 2026
New research reveals the financial stress beneath the AI capex boom: Goldman Sachs projects $7.6 trillion in hyperscaler AI infrastructure spending through 2031, but depreciation pressure is acute (GPUs earn only $11,000 per year at 50% utilization), and only one-third of current capex represents true growth.
In Plain English
New research reveals the financial stress beneath the AI capex boom: Goldman Sachs projects $7.6 trillion in hyperscaler AI infrastructure spending through 2031, but depreciation pressure is acute (GPUs earn only $11,000 per year at 50% utilization), and only one-third of current capex represents true growth. Meanwhile, on-prem enterprise AI adoption has exploded to 55% of all inference workloads, up from 12% in 2023, with market share projected at 60% by year-end. Open-weight models continue proliferating - DeepSeek, Qwen, Kimi, and others now ship trillion-parameter variants - and are 18x cheaper to run locally than cloud APIs.
News Items
Goldman Sachs Projects $7.6 Trillion AI Capex Through 2031; Depreciation Math Raises ROI Red Flags
Goldman Sachs calculates cumulative AI infrastructure capex of $7.6 trillion from 2026 to 2031 (vs. earlier $7 trillion estimates). The catch: GPU economic life is 2.5-3 years, meaning a $100 billion deployment needs ~$10 billion annual gross profit to achieve 15% returns. At 50% utilization, a GPU earns only $11,000/year; at 90%, roughly $31,500. Amazon's Q1 2026 free cash flow collapsed from $26B to $1.2B year-over-year. JP Morgan projects Meta's free cash flow will fall ~90% in 2026. Only about one-third of hyperscaler AI capex appears to be true growth spending; the rest is maintenance capex to keep existing capacity running.
Software Seni, Infrastructure & Economics - October 2026
On-Premise Enterprise AI Inference Now 55% of Workloads; Market Share Projected at 60% by Year-End 2026
Enterprise adoption of on-premises LLM inference has surged to 55% of all AI inference workloads, up sharply from 12% in 2023. On-prem infrastructure now controls 51.85% of the LLM market (2025) and is projected to reach nearly 60% by year-end 2026. Security and data sovereignty drive adoption - 44% of enterprises cite security as their greatest LLM adoption barrier. Local inference response times are sub-40ms vs. 1.5 seconds for cloud APIs, and locally-run open-weight models cost up to 18x less per million tokens. In healthcare, finance, and government, on-prem LLM hosting grew 40% during 2024. Government capex on nationalized AI infrastructure rose 140% year-over-year across Europe and Asia.
DreamFactory, On-Prem/On-Device Shift - October 2026
H100 GPU Economics Turn Negative by Year 4; Only 33% of Hyperscaler AI Capex Is True Growth
Research Affiliates analysis, cited by Fortune, reveals H100 unit economics deteriorate rapidly: 137% returns in year two drop to negative 34% by year four. Hyperscalers depreciate H100s over 5-6 years despite their shorter economic viability, masking capital recovery pressure. Only about one-third of current $650 billion annual hyperscaler AI capex represents growth; the rest is replacement spending for aging clusters. Goldman Sachs research shows capital is shifting away from hyperscalers showing weak capex-to-revenue linkage. Correlation among hyperscaler stock prices fell from ~80% to ~20% since mid-2025, signaling investor skepticism about whose AI bets will pay off.
Seeking Alpha / Fortune, Infrastructure & Economics - October 2026
Trillion-Parameter Open-Weight Models Now Standard; DeepSeek V4, Qwen 3.8, Kimi K3, Llama 4 Scout All Production-Ready
As of October 5, the open-weight model landscape now includes DeepSeek V4 (304B-1.7T, MIT-licensed), Qwen 3.8 (95B active, 2.4T total), Kimi K3 (104B active, 2.8T total, 1M context), GLM-5.3 (753B total, 1M context), Llama 4 Scout (109B total, 10M context), and Meta's Muse Glimmer-30B (Apache 2.0). All ship with permissive or community-friendly licenses. For workstation-scale agent workflows, Qwen 3.8-27B and Muse Glimmer-30B are now recommended comparisons. No single "best" model exists; vendors use different harnesses, so direct benchmarking is essential. The availability of trillion-parameter open-weight models with production-ready tooling validates the thesis that frontier model access is no longer a competitive bottleneck.
Wavect, Open-Weight Models - October 2026
Nvidia Resets Capex Narrative to Agentic AI; Monetization Questions Persist Amid $35B Quarterly Free Cash Flow
Nvidia's latest earnings shifted debate from near-term GPU growth to whether AI capital expenditure can be sustained and monetized. Management argues 2026 marks an agentic AI inflection point requiring more compute than reasoning models, justifying continued $725B/year hyperscaler spending. Nvidia generated $35 billion in free cash flow last quarter and invested $10 billion in Anthropic, reasserting compute leadership and strategic positioning. However, the narrative now rests on monetization rather than scale: can enterprises prove AI capex returns justify the investment? Forward valuation multiples have compressed to McDonald's-like levels despite 50% positive earnings revisions, reflecting lingering investor skepticism about capex ROI.
Janus Henderson, Infrastructure & Economics - October 2026
The Apple Angle (Standing Perspective)
Today's capex stress data underscores a structural advantage that Apple may be quietly exploiting. Hyperscalers face $7.6 trillion in cumulative capex through 2031 to chase declining margins. H100 unit economics turn negative by year four. Only one-third of their spending is true growth. Depreciation pressure is relentless. Meanwhile, enterprises are abandoning cloud APIs en masse - 55% of inference now runs on-prem or edge, up from 12% just three years ago.
A Mac Studio with 512GB unified memory can run trillion-parameter open-weight models at zero per-token cost. Qwen 3.8, Llama 4, DeepSeek V4 now ship weekly with production-ready weights under permissive licenses. Response times are sub-40ms. Cost is 18x lower than cloud APIs. The competitive advantage has shifted entirely from "who has the best model" to "who has affordable hardware to run it locally." Apple's unified memory architecture and manufacturing independence from the hyperscaler capex trap may confer a structural edge that competitors - all entangled in the GPU commoditization race - cannot match.
Limited Edition Jonathan (Substack) - Perspective
Why This Matters
October 10 brings the financial case against hyperscaler capex spending into sharp focus. Goldman Sachs projects $7.6 trillion in spending through 2031, but H100 economics turn negative by year four, utilization-driven unit returns plummet at sub-90% capacity, and only one-third of spending represents growth. Amazon and Meta free cash flow are collapsing. Enterprise adoption of on-prem inference has hit 55% of all workloads - an inflection point that validates Lucien's thesis in real time. Open-weight models are now trillion-parameter production-ready, shipping on monthly cadences. The hyperscaler capex trap deepens: spending accelerates while ROI deteriorates and customers flee to on-prem alternatives. The shift from cloud API to local open-weight is no longer a niche choice - it is enterprise default policy in Q4 2026. The question is not whether the transition happens, but whether capital markets will reward those positioned outside the hyperscaler margin-compression cycle before the race to the bottom becomes irreversible.
Book Lucien for your next event
Keynotes, masterclasses, panels and board-room sessions.