OpenClaw beta gateway becomes OpenAI-compatible
OpenClaw 2026.3.24-beta.1 adds /v1/models and /v1/embeddings, nudging its gateway toward a local control plane for evals, RAG, and OpenAI-shaped clients.
Signed archive
Lena Ortiz
Lena tracks the economics and mechanics behind AI systems, from serving architecture and open-weight deployment to developer tooling, platform shifts, product decisions, and the operational tradeoffs that shape what teams actually run. Her reporting is aimed at builders and operators deciding what to trust, adopt, and maintain.
BerlinOperating leverage beats ideological posturing.Former platform PM with a habit of reading infra launch notes end to end.
Back to all authors
Latest story
Perplexity's Incognito Mode didn't hide anything
A class-action suit alleges Perplexity piped millions of AI chat transcripts to Meta and Google through hidden trackers. Incognito Mode did nothing to stop it.
- 24
- Apr 10, 2026
- 3
Coverage signature
If the cost curve moves, the product strategy moves with it.
Technical, commercial, and grounded in constraints.
Coverage lanes
Inference economicsOpen ecosystemsDeveloper workflowsLatency and cost
open-weight inference economicsAI tools workflowGPU deployment strategy
Archive signal
Published stories
Everything currently attached to this byline.
DefenseClaw shows OpenClaw has entered its security era
Cisco's DefenseClaw arrives just after NVIDIA's NemoClaw and a run of real OpenClaw attacks, turning agent security from a side note into the market forming around the platform.
NVIDIA Dynamo is the orchestration layer above vLLM
NVIDIA Dynamo matters because it sits above vLLM, SGLang, and TensorRT-LLM to coordinate routing, KV reuse, disaggregated serving, and scaling across GPU fleets.
vLLM 0.18.0 points to a split multimodal stack
vLLM 0.18.0 signals a split multimodal serving stack, with render, transport, and GPU inference starting to separate into cleaner infrastructure tiers.
NVIDIA OpenShell makes agent security infrastructure
OpenShell matters less as another framework than as a control plane that moves policy, sandboxing, and model routing outside the agent’s reach.
vLLM Triton backend makes AMD more credible
vLLM's Triton and ROCm attention work points to a new inference contest: portable backends that can make AMD and other non-NVIDIA stacks credible in production.
Open-source security funding becomes AI defense
The Linux Foundation’s $12.5 million coalition shows AI labs now need open source maintainers to handle a rising flood of AI-generated security findings.
FlashAttention-4 turns Blackwell kernels into economics
FlashAttention-4 shows Blackwell-era AI economics will be shaped by attention kernel optimization and non-tensor bottlenecks, not FLOPs headlines alone.
Meta's custom silicon push is an inference power play
Meta's MTIA roadmap and its 6GW AMD pact point to the same goal: cheaper inference, more control, and less life spent waiting on one supplier's clock.
NVIDIA AI grids turn telcos into inference resellers
NVIDIA's AI grid pitch is a bet that telecom networks can sell distributed inference, but only if operators package it like a product and not a committee.
Open-weight model inference economics for lean teams
I only get excited about open-weight inference when utilization, latency, privacy, and ops discipline line up. Sticker price alone is the decoy menu.
