Today's News
The Pragmatic Engineer
Maude Lemaire, a principal engineer who spent nearly eight years leading Slack's backend performance team and recently tackled GitHub's database scaling challenges, has joined Cursor to work on performance improvements including time to first token. Her experience managing massive-scale systems—handling 500,000+ concurrent WebSocket connections and managing 16+ database shards—is directly applicable to the infrastructure demands of AI-native developer tools.
Techmeme
OpenAI released GPT-6 Cyber, a specialized model demonstrating concrete capability gains in AI-assisted security research. The model identified 81 previously unknown Firefox vulnerabilities and 948 Linux kernel exploits, with the research consuming $1.2M in API costs. This represents a new specialized model tier competing with general-purpose systems and signals OpenAI's strategy to develop domain-specific variants alongside flagship models.
Azeem Azhar, Exponential View
Azeem Azhar's infrastructure analysis reveals that actual 2027 data center supply will reach only 16.7–22.6GW, far below the speculative 50GW forecast circulating in markets. The critical constraint is power connection timelines, not physical construction: hyperscalers have already locked most 2027 power agreements while speculative shells face cancellation. The average US wait time for a power connection is four years, while data center shells take only 18–24 months to build. This creates a 20GW supply gap emerging in 2028 and gives communities and local governments genuine leverage to extract benefits before capacity becomes abundant.
Azhar notes that a 1GW data center with ~400,000 B300 chips costs ~$40B upfront but generates only ~$800M annual operating expenditure (roughly 2.5% of CapEx). Since each B300 chip generates ~$8/hour in revenue, every month of delay costs hyperscalers hundreds of millions. This asymmetry gives communities negotiating power. The North Lanarkshire deal in Scotland sets a benchmark: DataVita agreed to pay up to $100,000 per megawatt per year to a local fund—$50M annually at full 500MW capacity, or $543M over 15 years. By contrast, Oracle's New Mexico deal offered only $84,000 per megawatt as a one-time fee. At the North Lanarkshire benchmark, every GW of connection could yield $100M to local communities—only 0.2% of the $10–40B per GW that infrastructure companies can earn.
Clouded Judgement by Jamin Ball
Jamin Ball argues that the infrastructure stack for AI agents fundamentally differs from legacy cloud computing, which optimized for throughput across shared machines. Agent infrastructure instead requires isolated sandboxes per task, persistent state management, inference routing across models, and RL/training loops. This architectural shift suggests agent infrastructure could dwarf classical cloud application infrastructure in total addressable market, representing a structural shift in how compute will be provisioned and priced.
Lenny's Newsletter
As software increasingly operates autonomously, the new growth lever becomes getting users comfortable delegating work to the product through transparency, reliability, and control. Metrics like engagement may decline while actual customer value increases, requiring entirely different dashboards and KPIs than traditional product development.
The Pragmatic Engineer
The platform discovered that emoji reactions in large channels create exponential fan-out load—a 100,000-member channel with 100 custom emoji reactions generates 10 million fan-outs across the network. By capping custom emojis at 23, Slack prevents infrastructure collapse while maintaining user experience, illustrating how product constraints solve infrastructure problems at scale.
Bill Bishop at Sinocism
Three House Republicans are requesting closed-door testimony from Huang regarding his interactions with an Alice Huang at Megaspeed (a Nvidia subsidiary), who is publicly identified as a PRC intelligence asset, signaling potential congressional pressure on Nvidia over foreign espionage vulnerabilities.
Techmeme
responding to 43-1 passage of data center moratoriums across the country; this represents a significant policy concession on infrastructure transparency and grid management that could reshape how AI companies negotiate with utilities and communities.
Clouded Judgement by Jamin Ball
The acquisition signals validation of persistent filesystem infrastructure as a critical component of agent cloud stacks, where agents need to pause, snapshot, and fork state across task lifecycles—a capability distinct from traditional object storage.