AI is in full land‑grab mode: labs and investors are burning huge amounts of capital at the training layer while open, Chinese, and national models start to erode the big U.S. proprietary advantage. Governments are turning frontier models into regulated, sometimes national‑security infrastructure just as consumers and workers sour on the word 'AI' even while using the tools.
The real game this quarter is deciding which labs, regions, and tooling you let into your critical path before those choices harden into expensive dependencies.
Key Events
/OpenAI has racked up cumulative losses of $38.5B despite rapid revenue growth.
/Chinese lab DeepSeek raised $7.4B at a $50B valuation with a five‑year investor lock‑up.
/The U.S. government ordered Anthropic’s Fable 5 model shut down and restricted to U.S. nationals just days after launch.
/The EU AI Act high‑risk and transparency rules will start applying on August 2, 2026.
/Tesco is migrating 40,000 server workloads off VMware over Broadcom’s ‘abusive’ conduct.
Report
Capital is flooding into AI at oil‑and‑gas scale while most labs are still bleeding cash on training. At the same time, the stack is splintering by nation, regulation, and model type, which turns vendor choice into a balance‑sheet bet instead of a procurement detail.
the AI capex race is upside‑down on cash flow
OpenAI has reportedly lost $38.5B cumulatively while building its current model lineup. In 2025 alone it is said to generate $13B of revenue while still losing around $21B, even as it reports gross margins above 40% on model serving.
Industry‑wide AI capex is estimated in the hundreds of billions against only tens of billions of AI revenue, a mismatch that looks more like a land grab than normal enterprise software.
OpenAI is planning an 'AI research intern' built on roughly 500,000 A100‑equivalent GPUs, underscoring how capital‑intensive frontier training has become.
On the other side of the Pacific, Chinese lab DeepSeek just raised $7.4B at a $50B valuation with investors locked up for five years, signalling patient money is willing to underwrite similar loss curves.
geopolitics is rewriting the AI vendor map
The U.S. government ordered Anthropic’s Fable 5 shut down three days after launch and then restricted access for foreign nationals, moving a commercial frontier model into a national‑security regime overnight.
French president Macron personally lobbied to restore European access to Anthropic’s models at the G7, while at the same summit Dario Amodei and Demis Hassabis pushed for a U.S.-led AI coalition.
The EU AI Act will start enforcing high‑risk and transparency obligations from August 2, 2026, effectively forcing anyone doing serious AI in Europe into a regulated‑infrastructure posture.
France is already dropping Palantir for a domestic AI provider on core government data workloads, and the EU is working on its own models to reduce reliance on U.S. vendors.
At the same time, the U.S. has paused blacklisting of Chinese champion DeepSeek even as China designs a futures market for AI tokens and its securities regulator warns against AI‑driven speculation, showing how financial and security regulators are now entangled with model access.
open and Chinese models are commoditizing the base layer
Open‑weight model GLM‑5.2 now leads key benchmarks, scoring 51 on the Artificial Analysis Intelligence Index and matching Anthropic’s Claude Opus 4.8 while surpassing some GPT‑5.5 and Gemini 3.1 Pro evaluations.
A 4‑billion‑parameter open model has beaten some 30‑billion‑parameter models on web research benchmarks, reinforcing that efficiency and tuning can outweigh sheer size.
Chinese stacks like Qwen, Kimi K2.7, and DeepSeek V4 are now benchmark‑competitive and aggressively priced, with Kimi K2.7 reportedly at about $0.95 per million tokens.
ChatGPT’s market share has already dropped below 50%, landing around 46.4%, which means the default position of 'just use OpenAI' is weakening in practice.
New open automation frameworks such as Abacus AI’s Dragon models and n8n plus Claude pipelines are being used to replace traditional SaaS workflows, cutting into the idea that proprietary APIs alone constitute a durable moat.
AI is widely used but deeply disliked
About 57% of Americans say they have used AI tools, and 42% report productivity gains from them. At the same time, only 16% of Americans think AI will have a positive impact on society, and 60% say that 'AI' in brand messaging is a turnoff.
Roughly 1 million IT jobs in the U.S. have been lost since the rise of AI technologies, feeding the perception that AI primarily benefits a small elite.
Inside big tech, Meta has laid off about 8,000 people and its CTO says morale is near the worst ever, while workers increasingly resist AI‑heavy initiatives like internal hackathons.
Even upbeat narratives like Jeff Bezos predicting AI‑driven labor shortages rather than job losses are met with skepticism, as employees see companies citing AI to justify cuts in real time.
infrastructure and tooling are being renegotiated in real time
Tesco is migrating around 40,000 server workloads off VMware, explicitly blaming Broadcom’s 'abusive' conduct, and other enterprises report that similarly large moves are painful but tractable.
Alternatives like HPE’s Morpheus and Red Hat OpenShift are gaining traction, with HPE even offering a year of free software to lure VMware customers, while users complain that the broader VMware ecosystem is deteriorating.
On the tooling side, Y Combinator’s Paxel was found to send Cloudflare OAuth tokens and git emails from developer machines to YC servers via a one‑line install, confirmed through live HTTPS interception.
The Pentagon now has about 1.5 million personnel using generative AI tools, and observers worry that sensitive internal data could be misused to train private models without adequate safeguards.
In parallel, millions of copyrighted songs have been quietly fed into AI music generators while labs refuse to disclose training data volumes for fear of lawsuits, expanding the IP risk surface around seemingly innocuous AI features.
What This Means
The AI boom is shifting from a single dominant platform story to a messy contest between capital‑intensive frontier labs, cheaper open and national models, and increasingly hostile publics and regulators. The live decision is which dependencies—labs, regions, and tooling—you are willing to let onto your critical path knowing that switching later will be expensive, public, and possibly political.
On Watch
/Uber’s plan to launch a premium robotaxi service in Houston by 2027 will test whether autonomous ride‑hailing can move beyond limited pilots into a scaled, paying market.
/China’s work on a futures market for AI tokens hints at full financialization of model access and could pull securities regulators deeper into AI governance.
/An emerging ecosystem of EU AI Act compliance tools within the open‑source community may decide which vendors are actually sellable into Europe once high‑risk obligations bite in 2026.
Interesting
/- China is offering free AI technology to developing nations, contrasting with G7 discussions on access to American AI models.
/- The CEO of DeepSeek, Liang, invested $3 billion of his own funds, becoming the largest stakeholder in the recent funding round.
/- Concerns about OpenAI's reliance on government contracts for financial stability have been raised, questioning its long-term viability.
/- The Pentagon's use of Grok AI for missile strikes has been described as a controversial move, raising questions about the implications of AI in warfare.
/- Predictive AI agents are expected to replace traditional BI dashboards by 2027, offering explanations and recommendations.
We processed 10,000+ comments and posts to generate this report.
AI-generated content. Verify critical information independently.
/OpenAI has racked up cumulative losses of $38.5B despite rapid revenue growth.
/Chinese lab DeepSeek raised $7.4B at a $50B valuation with a five‑year investor lock‑up.
/The U.S. government ordered Anthropic’s Fable 5 model shut down and restricted to U.S. nationals just days after launch.
/The EU AI Act high‑risk and transparency rules will start applying on August 2, 2026.
/Tesco is migrating 40,000 server workloads off VMware over Broadcom’s ‘abusive’ conduct.
On Watch
/Uber’s plan to launch a premium robotaxi service in Houston by 2027 will test whether autonomous ride‑hailing can move beyond limited pilots into a scaled, paying market.
/China’s work on a futures market for AI tokens hints at full financialization of model access and could pull securities regulators deeper into AI governance.
/An emerging ecosystem of EU AI Act compliance tools within the open‑source community may decide which vendors are actually sellable into Europe once high‑risk obligations bite in 2026.
Interesting
/- China is offering free AI technology to developing nations, contrasting with G7 discussions on access to American AI models.
/- The CEO of DeepSeek, Liang, invested $3 billion of his own funds, becoming the largest stakeholder in the recent funding round.
/- Concerns about OpenAI's reliance on government contracts for financial stability have been raised, questioning its long-term viability.
/- The Pentagon's use of Grok AI for missile strikes has been described as a controversial move, raising questions about the implications of AI in warfare.
/- Predictive AI agents are expected to replace traditional BI dashboards by 2027, offering explanations and recommendations.