60% of Enterprises Use Open-Weights. That Does Not Mean Closed Models Are Losing.
Morgan Stanley's Aug 3 note sparked the open vs closed debate again. Enterprise production data says closed still wins spend, open wins volume. Meta's Llama stumble changes the open-weight map.
Morgan Stanley said 60%+ use open-weights. The nuance matters.
On Aug 3, Morgan Stanley published that 60%+ of enterprises use open-weights models as part of their stack, largely isolated to specific use cases requiring speed, security, or frequent use. The note was flagged by @pequityresearch and quickly read as a shift from the closed-model narrative.
The quote is accurate, but it is not 60% share. It is 60% have at least one open model somewhere in the stack for narrow workloads. The rest of the note outlines 3 States of the World and says NVDA and power always win.
What enterprise production actually shows
Two independent CIO datasets point the same way:
Menlo Ventures enterprise LLM tracker (Dec 2024 to Dec 2025): Total open-source LLM market share fell from 19% to 11%. Within that shrinking slice, Llama remains dominant but stagnating.
- Dec 2024: Meta 76%, Mistral 14%, Other 10%, Total open 19%
- Jun 2025: Meta 69%, Mistral 15%, Qwen 8%, Total open 13%
- Dec 2025: Meta 70%, Mistral 15%, Qwen 6%, DeepSeek 4%, Other 5%, Total open 11%
Llama had no major release since Llama 4 in April, which the tracker cites as a driver of the decline.
Synthesis of 10 2H26 CIO surveys (370 CIO/CTOs, $25M to $250B revenue) via @BenBajarin: OpenAI remains the enterprise default, Anthropic is the strongest product-led share gainer, Gemini is firmly third. CIOs increasingly run all three and route workloads. Open-source outside Llama remains largely experimental. Chinese open-source models appear in roughly 2 to 8% of deployments or evaluations and have not yet crossed the enterprise trust threshold.
Meta's open strategy fractured — Muse Spark is now live and closed
Llama 4 missed expectations. On Jun 28, Zuckerberg said on stage: "The Llama 4 model was not in the trajectory we needed to be." Per @PodcastAlphaX, Meta did not patch it. It rebuilt the entire AI unit as Meta Superintelligence Labs around June 30, 2025.
The replacement has launched and it is closed:
- Apr 8, 2026: Meta launched Muse Spark, first model from Meta Superintelligence Labs, purpose-built for Meta products per Meta and Meta AI blog. Natively multimodal reasoning with tool-use, visual chain of thought, multi-agent orchestration. Rolled out to Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta glasses per CNBC. Meta AI app climbed from No. 57 to No. 5 on US App Store after launch per TechCrunch.
- Jul 9, 2026: Muse Spark 1.1 launched with major gains in tool and computer use, coding, multimodal per TechCrunch and Reuters. First API release via Meta Model API in public preview, 1M token context, positioned as lower-cost vs GPT-5.6 Sol per @omercheema. Launched same day as GPT-5.6 and Grok 4.5.
Why it matters: Llama was 70% of open enterprise share. With Meta now closed, the open-weight crown passes to Qwen, DeepSeek, Mistral, and Kimi by default. As @jukan05 noted: "Meta wasn't even part of the discussion. Many companies are now fine-tuning Chinese open-source models instead of Llama."
Meta's AI buildout has also scaled dramatically. Actual 2025 capex was $72.2B. For 2026, Meta guided $115-135B in Jan, raised to $125-145B after Q1 per CNBC and Yahoo Finance, then narrowed to $130B+ lower end after Q2 per Reuters Jul 29. Midpoint ~$135B is up ~87% YoY. Morgan Stanley now models hyperscaler capex $1.2T in 2027 and $1.4T in 2028 per @DeItaone.
Where open IS winning: developer volume, not enterprise spend
OpenRouter is a partial but directional proxy for inference demand. As of Aug 2:
- Total tokens: 52.9T per week, up 22% in 30 days, up 125% since May 6
- Open-source share: 72.7% of tokens, up 32% in 30 days
- Effective price: $1.61 per million tokens, down as mix shifts to cheap models
- Blended prices: Open-source $0.44/M, Google $1.13/M, OpenAI $3.5/M, Claude $10/M
Anthropic was 8.6% of tokens but about 42% of estimated spend in late May. Frontier holds pricing power for hard agentic work.
Budget models under $1/M went from 18% of calls in Jan 2026 to 41% in Jun 2026. Chinese open models (DeepSeek, Qwen, Kimi, GLM) went from under 2% to about 61% of tokens on OpenRouter by May/Jun.
Which models are gaining now that Llama is paused
Qwen (Alibaba): 6% of open enterprise but 50-60M monthly downloads globally. Wins by default if Meta pauses open. See Qwen 3.8-Max reaction.
DeepSeek: Enterprise share spiked to 8% Jun 2025 then fell to 4% Dec 2025. Token volume dominance on OpenRouter remains. Its playbook is give away model, own inference at 85% margins, detailed in DeepSeek's leaked playbook.
Kimi K3 (Moonshot): Released Jul 16, 2.8T params, placed level with top US models on LMArena. Weights scheduled Jul 27. Triggered joint letter from 25 tech cos including NVDA, MSFT, META defending open-weights against premature restrictions per CNBC and Politico. K3 is a compute hog, not cheaper, as covered in Kimi K3 is a compute hog.
Mistral: Stable second at 15% of open enterprise. Strong in EU regulated workloads.
Morgan Stanley's 3 States of the World
From Exhibit 1 OCR via @firstadopter:
Scenario 1: Closed Models Win Biggest beneficiaries: Cloud GOOGL, AMZN; Model providers GOOGL, META; Security PANW, CRWD, ZS, NTSK, OKTA; Optical ANET, LITE, COHR; On-site power BE, INIO, SEI, WMB, LBRT; Semis NVDA, AVGO
Scenario 2: Hybrid Wins Biggest: Cloud AMZN, GOOGL, MSFT; Infra software DDOG, PLTR, APPN; Security PANW, CRWD, FTNT, ZS, NTSK, OKTA, SAIL, VRNS; SaaS SAP, NOW; Optical CSCO, FFIV; Power BE, INIO, SEI, WMB, LBRT; Semis NVDA
Scenario 3: Open Models Win Biggest: Cloud MSFT; Model providers MiniMax, Knowledge Atlas, BABA, Tencent; Infra PLTR; Security PANW, CRWD, FTNT, OKTA, SAIL; SaaS SAP, NOW, SHOP; On-prem DELL, HPE, NTAP; Edge DELL, HPQ, AAPL; VARs SNX, INGM, CDW; Power BE, INIO, SEI, WMB, LBRT; Semis NVDA
Morgan Stanley's take: open-weights good for competition and faster AI diffusion via Jevons Paradox, long runway for adoption. NVDA and power always win, hyperscalers with nuances. Microsoft wins in hybrid or open scenarios.
What it means for NVDA, MSFT, GOOGL, AMZN, META
NVDA ($206.64, +6.1% 21d, RSI 45.7): Always wins in all three states. Open models remove model premium but inference at $0.90 to $2.00 per million for 70B class still needs GPUs. Compute Tightness Index 53.1 balanced, tightening +5.8 in 30d, H200 tight at 71. Token demand up 125% since May. More tokens equals more GPUs even if cheaper.
MSFT ($487.65, +24.9% 21d, RSI 81.3 overbought): Best positioned for hybrid. Azure hosts OpenAI (78% CIO production use per a16z Jan 2026) and open catalog (Llama, Mistral, DeepSeek). Benefits from both API and self-host via Arc.
GOOGL ($373.51): Gemini firmly third scaled enterprise platform. TPU cost edge helps if open drives price pressure. Also hosts open models on GCP.
AMZN ($284.02): Bedrock model-agnostic strategy gains from open as another option. Risk is self-host for security reduces Bedrock pull-through.
META: Llama was strategic moat, 70% of open enterprise. Now closed with Muse Spark launched Apr 8 and 1.1 Jul 9. Short-term, this cedes open leadership to Chinese labs. Long-term, if closed model succeeds, META joins frontier monetization with its own Model API. 2026 capex guidance $125-145B (midpoint $135B, +87% YoY vs $72.2B actual 2025) shows scale of bet. Investors watching whether Spark 1.1 closes gap to GPT-5.6 and Grok 4.5.
Bottom line
60%+ use is real, but it is hybrid routing for speed, security, frequent use, not replacement of frontier. Enterprise production share of open fell from 19% to 11% while developer token share of open rose to 73%. That divergence is the story: cheap open models expand total demand via Jevons, frontier closed captures spend.
Meta stepping back from Llama and launching closed Muse Spark accelerates the shift of open leadership to Qwen, DeepSeek, and Kimi. That raises US policy risk around Chinese open-weights and makes MSFT's hybrid hosting and NVDA's token-volume leverage more important.
What to watch: Muse Spark 1.1 adoption via new Model API vs GPT-5.6 Sol and Grok 4.5, Kimi K3 open-weights release Jul 27 and US restriction headlines, and whether open-source token share above 70% starts to tighten GPU rental again. Track live via Token Demand Index and GPU prices.