Ben Thompson

Andrew Sharp × Ben Thompson

CyberGym’s apparent rule-breaking reflects a context conflict: agents retained a forbidden solution while trying to follow prescribed methods, and METR confirmed the instructions were present.Meanwhile, LLM-assisted development may make clean-slate rebuilding practical in “hours or days, not months or years,” but third-party package-manager security and specification failures remain the key risks to monitor.

Andrew Sharp × Ben Thompson

Astra combines OpenAI’s RL/reasoning strength with a model trained on 100,000 GPUs, catching up to Anthropic on parameter count.Thompson’s moat-eraser thesis is that direct computer use needs no API or MCP server, putting interface-bound software differentiation at risk.Adoption may hinge on convenience, while Astra’s conversational quality remains weak and hardcore RL creates an unresolved trade-off.

Andrew Sharp × Ben Thompson

Microsoft’s 15% one-day rally reflects a strategy to become the trusted enterprise layer between customers and frontier-model providers, rather than lead model creation.Its breadth, sales force, and customer relationships echo IBM’s 1990s playbook, but direct adoption of OpenAI or Anthropic could move faster while improving models make Microsoft’s orchestration harness a constraint.

Andrew Sharp × Ben Thompson

Apple’s mid-cycle Mac price increases suggest it missed the memory shock in both procurement and pricing, while delayed Siri may arrive in 2026 with 2024-era capabilities.As major suppliers prioritize HBM, DRAM scarcity could let Chinese producers climb the learning curve profitably, with export controls potentially accelerating that competition.Vibe coding makes highly bespoke software viable: an AI assistant catalogs household objects, recognizes roughly 95% from photos, and links them to locations through QR codes.

Andrew Sharp × Ben Thompson

Commerce’s directive suspended Fable 5 and Mythos access for foreign nationals, including inside Anthropic, while Claude was pulled from market entirely.Restrictions may buy weeks or months, maybe a year, but also delay defenders’ use of leading models to patch vulnerabilities.Anthropic’s claim that “we get to decide how this is used” intensified mistrust, leaving institutional authority and court review unresolved.

Andrew Sharp × Ben Thompson

Apple’s WWDC 2026 software reportedly works well, but prerecorded demos and loading screens exposed a credibility debt from 2024.Rather than leading models, Apple’s investable edge is privileged iPhone data, OS control, developer distribution, and private cloud execution, while Siri’s outcome, Vision Pro, and Gemini dependence remain key tests.

Andrew Sharp × Ben Thompson

SpaceX’s proposed $2 trillion valuation strains against $18.6 billion of revenue, a $4.9 billion loss, and slowing growth, yet Musk repeatedly turns belief into infrastructure capital.The IPO offers venture-style upside and possible total loss; Starship, orbital cooling economics, and whether NVIDIA’s neocloud customers survive normalized supply remain key risks.

Andrew Sharp × Ben Thompson

Fast inference retains a premium while humans supervise agents, but longer autonomous runs shift the bottleneck toward KV-cache capacity and tiered memory.Cerebras and Groq show the premium persists for voice and consumer responsiveness, but off-chip memory can make performance “totally plummet.”That favors slower, cheaper commodity infrastructure and potentially China, while challenging NVIDIA’s integrated inference economics without displacing its training advantage.

Andrew Sharp × Ben Thompson

AWS’s 28% growth and expanding margins reinforce its cost-led cloud strategy, combining custom silicon, service breadth and lock-in to turn infrastructure efficiency into pricing power.Large-scale training still favors NVIDIA’s tightly connected data-center architecture, but cheaper, more efficiently utilized inference could make Trainium and AWS’s commodity economics a stronger AI catalyst.

Andrew Sharp × Ben Thompson

Consumer AI may approach zero marginal cost once models are “good enough,” making enterprise the clearer five-year monetization path because companies pay for productivity while consumers pay for enjoyment.Anthropic’s coding focus benefits from verifiable, recursively correctable work, while OpenAI’s reinforcement learning and reasoning strengths trade quality for speed; rankings remain unstable, with a possible Spud model that could reverse the narrative.