OpenAI appointed Jill Hazelbaker chief corporate affairs officer, putting her in charge of global communications and corporate positioning. She joins from Uber, where she spent a decade as president and chief corporate affairs officer, overseeing global communications, marketing, public policy, safety and people teams.
At Uber, Hazelbaker helped navigate regulatory and reputational challenges as the company grew from an ambitious startup into one of the world's largest global technology platforms. She previously led communications and public policy at Snap and held senior roles at Google covering global corporate communications, communications across Europe, the Middle East and Africa, and government relations in Europe.
Hazelbaker also served as national communications director and chief spokesperson for John McCain's 2008 presidential campaign.
Why it matters: OpenAI can now draw on Hazelbaker's experience navigating Uber's regulatory and reputational challenges as it engages the people and institutions shaping AI.
Anthropic’s compute bill consumed 58% of revenue in the quarter ended September 30, up from 48% the previous quarter. Quarterly revenue rose to $17 billion from $11.5 billion, but the operating loss also grew, according to an update to the company’s IPO filing.
Compute expenses reached $10 billion, including $5 billion bought through short-term commitments that cost more than comparable multiyear contracts. Revenue per dollar of compute peaked at just over 2x before falling to 1.72x.
To meet demand, Anthropic turned to higher-priced neocloud capacity from SpaceX, Nscale, Akamai, and Volta. Its Anthropic Direct project aims to build more of its own compute base, but the filing says infrastructure will remain its largest operating expense for the foreseeable future.
Why it matters: For Anthropic’s IPO investors, the risk is that demand keeps forcing expensive capacity purchases: the company says compute and infrastructure costs will rise in absolute dollars even as its $10 billion quarterly compute bill already outpaces the previous quarter’s margin profile.
OpenAI’s hardware team used GPT-6 Astra to design a device that lets users talk to the Codex coding agent in under a week. According to OpenAI, Astra handled the enclosure, PCB layout, component selection, trace routing, firmware, and user interface.
A contract manufacturer produced a small prototype batch in under a month, but several revision rounds followed because Astra missed manufacturing constraints, tolerances, and assembly details. The result is an accelerated design workflow—not an autonomous hardware pipeline.
OpenAI also says a newer undisclosed internal model scored 0.875 on BenchCAD 2.0, versus Astra’s 0.594 at its highest reasoning level. The benchmark tests 3D-part and assembly reconstruction plus PCB-connection identification.
Why it matters: For OpenAI’s hardware team, Astra cuts initial device design to under a week, but the required manufacturing revisions show that contract manufacturers still handle the gap between model-generated designs and buildable hardware.
Anthropic’s new Message Threads feature cuts time to first token by about 35% for Claude requests containing 500,000 to one million tokens. Released as a public beta, it stops API clients from resending the entire conversation after every message, transmitting only what changed instead.
The feature targets coding agents, computer-use agents, and workflows processing large volumes of documents or images. Message Threads works on Anthropic’s Claude platform and through AWS, Microsoft Foundry, and Google Cloud.
The speedup does not reduce token consumption: Claude still reads the full conversation history. Threads remain encrypted in memory and expire after five minutes of inactivity, limiting their value for intermittent workloads.
Why it matters: For teams building coding and computer-use agents on AWS, Microsoft Foundry, or Google Cloud, Message Threads makes long sessions respond faster but offers little help to intermittent workloads that outlast its five-minute expiration window.
An LLM serving team can use a six-question decision tree to match bottlenecks with fixes: prefix caching for identical prompts, continuous batching for mixed-length generations, PagedAttention for wasted KV-cache memory, speculative decoding for per-user streaming latency, and prefill-decode disaggregation when long prompts delay generation. Model fit comes first: a model that barely fits leaves little room for KV cache, limiting batch size. Most modern serving engines already support continuous batching and PagedAttention, so teams should check the bottleneck before adding configuration.
AI search works well for common or complex queries with abundant online information, but traditional Google is better for precise specifications, breaking news, and health, legal, or financial questions. Stanford research found extraction works well once AI finds an accurate source; UC San Diego researchers found systems may favor relevance over authority.
Why it matters: For LLM serving engineers, matching the fix to the bottleneck avoids unnecessary serving complexity; for users checking medical, legal, financial, or breaking-news claims, Google keeps source choice more visible than an AI-generated answer.
Fal raised $1.3 billion at a $16 billion valuation, doubling its $8 billion valuation from April. Bessemer Venture Partners led the round, with Nvidia, Sound Ventures, and T. Rowe Price participating. The San Francisco startup hosts and runs video, image, and audio AI models.
Fal announced its Series E and F rounds together, bringing combined proceeds to $1.65 billion; Sequoia led the earlier round. The Series F arrives as demand rises for inference services, particularly from customers using open-source models. Fal surpassed $800 million in annualized revenue earlier this year.
Baseten, Fireworks AI, and Modal are also raising large rounds at higher valuations. But high compute costs keep inference companies’ gross margins thin, according to people familiar with their finances.
Why it matters: Fal has shown that inference demand can support a $16 billion valuation before margins are fixed; Baseten, Fireworks AI, and Modal now face the same expensive-compute test.