Some startups, like Harvey, Abridge, Ramp, and Rogo, are embracing open-weight models or training their own models to reduce expensive reliance on frontier labs
THE SO WHAT
When a $15.6 billion legal startup like Harvey starts leaning into open-weight or self-trained models, the cost and control calculus around frontier APIs is shifting. If your gross margins depend on third-party model pricing, you should be running serious build-or-borrow analysis on domain-specific open models this quarter.
READ THE SOURCE
MORE FROM THE WIRE
Meta's Muse gives Zuckerberg another shot at building a real platform
If Muse becomes the default agent layer across Meta’s surfaces, the real platform isn’t the model—it’s the distribution and identity graph wrapped around it. For builders, the question is whether you treat Meta as an acquisition channel for agents or as the primary runtime for your workflows.
Applied AIAmazon and Shopify make starkly different moves in the brewing battle over AI shopping
AI shopping is fragmenting into closed ecosystems and interoperable rails — Amazon is tightening control over the assistant layer while Shopify leans into partnering with agents like Muse on payments. If you own a commerce surface, decide now whether you’re a destination with your own AI or an infrastructure layer for third-party agents.
Applied AIOpenAI forms math advisory group as its AI resolves more than 100 open problems
If models are credibly resolving 100+ open math problems, the bottleneck shifts from capability to governance — hence an advisory group that can’t actually slow the work. For R&D-heavy orgs, the lesson is clear: you need domain-specific review and publication policies ready before your internal models start producing results your existing process can’t safely absorb.
Applied AI1.58-million-staff Amazon will soon have more Nvidia GPUs than employees — AWS to buy more than 3 million additional chips before 2029 in addition to thousands of existing H100, H200
Ordering more than 3 million additional Nvidia GPUs — on top of near-fully-subscribed Trainium3 — is Amazon treating AI compute as a multi-year utility buildout, not a discretionary cloud SKU. If you’re betting on at-scale training or inference, assume hyperscaler capacity will exist but economics and prioritization will be the constraint you negotiate around.