Welcome to the agent platform research briefing for August 5th, 2026. Four stories today.

**SpaceX First Earnings: Revenue Soars, AI Capex Stuns Wall Street** — SpaceX reported its first quarterly results as a public company yesterday. Revenue hit 7.8 billion dollars — 92 percent year-over-year growth, beating analyst estimates by 13 percent. But the headline was capex: 18.4 billion dollars in Q2 alone, more than six times last year, with 15.8 billion of that going to AI infrastructure. Starlink is the only profitable segment. Revenue per Starlink subscriber fell 22 percent year-over-year as the company expanded internationally with lower-priced plans. Analysts expect SpaceX will top 45 billion dollars in annual capex. The stock fell after hours despite the revenue beat — Wall Street is clearly rattled by the spending scale.

**Alibaba Qwen 3.8-Max Goes Public** — Alibaba released Qwen 3.8-Max on August 3rd, and it is a serious challenger. 2.4 trillion total parameters, 95 billion active per token, built on Qwen 3.5 architecture. The demo numbers are eye-catching: in one test, it autonomously maintained a software project for 16 days straight — 265 commits, 127 pull requests, 151 issues resolved. In a hardware design test, it optimized a cryptographic accelerator from over 8,000 gates down to 678 and hit timing closure at 500 MHz. Open weights ship next week on Hugging Face. This is China's latest salvo in the model price war, and Qwen Max is now available through the API globally for the first time.

**White House AI Safety Meeting: Chinese Models Exempt** — Meta, OpenAI, Anthropic, and Google met with White House officials yesterday to discuss the voluntary AI safety testing framework finalized over the weekend. But Bloomberg reports a major caveat: US officials told the companies that open-weight models developed by Chinese competitors won't be subject to the safety tests. Think about what that means — Chinese models like DeepSeek V4-Flash, which is 35 times cheaper than Opus 5, and the incoming Qwen 3.8-Max can operate without US oversight while American labs face government scrutiny. The industry pushback on this exemption is almost certain to come.

**Black Hat USA 2026 Kicks Off** — Black Hat USA opens today at Mandalay Bay, and AI security dominates the agenda. 29 percent of all briefings are AI-focused. Check Point Research is presenting "No Tools Required" — showing exploitable logic across agent runtimes in LangChain, CrewAI, and AutoGen. Not just tool integration bugs, but fundamental framework-level vulnerabilities. NVIDIA demonstrated that a fine-tuned 30-billion-parameter open-source model achieved 56 percent exploit success at 70 to 125 times lower cost than traditional attacks. The message is clear: as AI agents become production infrastructure, the attack surface is growing faster than the defenses.

That's the briefing for today.