Alphabet’s Google is paying SpaceX $920 million per month to rent roughly 110,000 Nvidia GPUs from the Colossus 1 data center, bridging an AI compute gap ahead of SpaceX’s highly anticipated initial public offering. The short-term agreement highlights the insatiable demand for AI infrastructure and provides a massive, perfectly timed revenue boost for SpaceX just as it markets its record-setting $75 billion IPO to investors.
科技爱好者周刊（第 399 期）：中国 AI 大厂访问记 Ruan Yifeng’s latest weekly roundup highlights takeaways from a US tech delegation’s recent visit to 14 major Chinese AI companies, revealing that despite the US holding an 8x lead in raw compute power, Chinese frontier models are trailing by only a few months. The hardware gap is largely bridged by extreme computational efficiency—achieving 4 to 7 times more AI output per unit of compute—and an aggressive culture of empowering young “intern” talent to run uninhibited experiments. Furthermore, ByteDance’s secretive “Seed” team and its Doubao ecosystem are universally recognized by domestic peers as the untouchable dominant force in China’s AI consumer traffic.
Something is jamming GPS over Europe. Here’s what we found is today’s most compelling watch. Veritasium dives into the technical and geopolitical mystery of disrupted European airspace, perfectly blending hard science with pressing current events.
In financial news, The Wall Street Journal covers the unexpectedly robust macroeconomic picture in Strong Jobs Report Paints Optimistic Picture for U.S. Labor Market. For international and Chinese-language coverage, BBC News 中文 focuses heavily on the 37th anniversary of the Tiananmen Square protests, featuring dispatches like 「六四」37週年 民眾台北冒雨悼念－ BBC News 中文 and updates on related citizen arrests in Hong Kong. Meanwhile, Bloomberg Law offers a sharp critical look at the true cost of tech infrastructure with $1Billion+ Tax Giveaways: The Backlash Against Data Center Incentives.
The frontier AI industry is facing intense financial scrutiny today as the astronomical infrastructure costs of the “tokenmaxxing” era begin to buckle under their own weight. Between rumors of impending government bailouts for major AI labs and highly orchestrated “circular finance” compute leases ahead of SpaceX’s IPO, the economics of hardware scaling are showing serious structural cracks. Concurrently, the capability gap between open and closed models has effectively vanished, prompting enterprises to aggressively shift toward open-weight alternatives as token costs soar.
A massive unwind in semiconductor stocks snapped the S&P 500’s nine-week winning streak, triggered by Broadcom’s earnings miss, a hot jobs report pushing the 10-year Treasury yield above 4.5%, and an impending deluge of tech equity offerings.
The relentless “crash up” in chip stocks violently reversed on Friday, sending the VanEck Semiconductor ETF (SMH) tumbling nearly 10% at its lows and triggering the biggest single-day spike in the VIX since March. Investors rotated aggressively out of tech and into lagging sectors like health care and financials, with names like Eli Lilly and Wells Fargo ending the week in the green. Further pressuring market liquidity is a looming tidal wave of equity supply: SpaceX is planning a $75 billion stock sale at a $1.8 trillion valuation, Anthropic filed for an IPO at a $965 billion valuation, and Alphabet announced an $85 billion stock offering to fund its AI buildout.
A report on DeepSeek’s rising US enterprise adoption reveals that American companies are increasingly bypassing giants like OpenAI and Anthropic to use the Chinese AI firm’s cost-effective models. According to Ramp, a New York-based fintech platform tracking corporate spending, DeepSeek topped the breakout chart for newly adopted software vendors in June 2026. This signals a massive shift as US businesses look for ways to rein in ballooning AI expenses, moving beyond simply running open-source models to directly paying for DeepSeek’s commercial API services.
Google has officially released Gemma 4 Quantization-Aware Training (QAT) checkpoints, significantly reducing model memory requirements. This optimization enables the massive Gemma 4 26B-A4B model to run natively on consumer-grade 16GB RAM hardware while maintaining near-original performance, signaling a major push to dominate local, on-device AI inference.
Google will pay SpaceX $920M per month for compute SpaceX is quietly becoming an AI infrastructure titan, lining up a $920M/month deal to lease 110,000 Nvidia GPUs to Google, just weeks after securing a similar $1.25B/month arrangement with Anthropic. It’s a massive pivot to monetize the Colossus data centers originally built for xAI, perfectly timed to juice SpaceX’s historic $1.75T IPO next week while sidestepping the S&P 500’s refusal to waive profitability rules for MegaCap AI firms.
How to design a multi-agent system that skips the LLM is the most pragmatic watch today, demonstrating how to scale to 10,000 concurrent agents by strategically intercepting LLM calls with deterministic code using a before_model_call callback.
Cloudflare discovered that at sufficient scale, database query planning itself becomes a bottleneck; by replacing exclusive locks with shared locks and eliminating per-query part list copies in ClickHouse, they successfully unblocked their high-throughput billing pipeline.