The AI Brief

Oracle shed 21,000 jobs, almost 13% of its workforce, in the past year. The company's total workforce stands at 141,000 full-time employees as of May 2026, disclosed in its annual regulatory filing. Oracle stated in the filing: "The adoption and deployment of AI technologies across our operations have resulted, and may continue to result, in reductions to our workforce." The statement is notable both for its candor and its forward guidance: Oracle explicitly reserved the right to cut further.

The company spent $1.8 billion on restructuring costs, including severance payments and other exit costs, a jump from the $374 million it spent on restructuring the previous year. Oracle's sales and marketing workforce saw one of the sharpest falls, dropping from approximately 31,000 employees to 25,000, a reduction of around 6,000, or roughly 19 percent. In its fiscal 2026, Oracle spent $55.7 billion on capital expenditures, a 162% increase from the $21.2 billion it spent in fiscal 2025. The pattern is textbook: capex for AI infrastructure rises sharply while headcount falls and the connection is made explicit.

Oracle's free cash flow plummeted to negative $23.7 billion. Yet the company reported remaining performance obligations worth $638 billion, up from $138 billion last year. Oracle holds a five-year, $300 billion deal to provide data center capacity to OpenAI, one of its largest AI agreements. The disclosure creates a legal template: if Oracle's language triggers no adverse regulatory reaction, expect comparable language to migrate into other large enterprise filings across 2026.

Baseten announced a $1.5 billion Series F financing led by Altimeter Capital, Conviction, and Spark Capital on June 22, 2026. The round includes investments made across two tranches at $13 billion and $11 billion respectively, and reflects surging demand for inference at the app layer as closed-source and open-source models converge in capability, cost, and customization.

Just five months ago, the startup announced that it had raised a $300 million Series E at a $5 billion valuation. If finalized, this latest round would represent a 160% increase in valuation in less than half a year. At the end of the first quarter, Baseten's annualized revenue came to around $600 million, compared to $200 million at the beginning of the quarter. The growth was attributed to an explosion of apps using open-source AI models. Customers named in the announcement include Cursor, Clay, and Abridge. The Fable 5 export suspension has accelerated enterprise interest in open-source inference alternatives, providing structural tailwind for independent inference providers.

DeepSeek-V4 Preview is officially live and open-sourced. Both models support 1M context and dual modes (Thinking and Non-Thinking). DeepSeek-V4-Pro has 1.6 trillion total parameters with 49 billion active per forward pass. Both models support 1M context with dual modes. The legacy deepseek-chat and deepseek-reasoner endpoints will be fully retired and inaccessible after July 24, 2026, 15:59 UTC.

The migration is not a like-for-like swap. During the grace period, deepseek-chat routes to V4-Flash in non-thinking mode and deepseek-reasoner routes to V4-Flash in thinking mode. Neither routes to V4-Pro, so moving to deepseek-v4-pro is an upgrade, not a like-for-like swap. In the 1M-token context setting, DeepSeek-V4-Pro requires only 27% of single-token inference FLOPs and 10% of the KV cache compared with DeepSeek-V3.2. That efficiency gain is material for high-throughput pipelines.

On the competitive benchmark picture: the independently tracked number is DeepSeek-V4-Pro-Max at 80.6% on SWE-bench Verified, the highest open-weights entry, tied with Gemini 3.1 Pro. Closed frontier models score higher: Claude Fable 5 leads at 95.0%, though it is currently suspended. Scale's SEAL leaderboard has no DeepSeek V4 result, so agentic performance under a standardized harness is unverified. Teams currently routing through the legacy aliases via OpenRouter or third-party wrappers should verify whether upstream providers have already migrated transparently or will break on the deadline.

A24 and Google DeepMind announced their joint venture on June 22, 2026. The tech giant will fund an AI research lab at A24 to the tune of $75 million to make tools available to filmmakers. The partnership will give A24 access to DeepMind's research and infrastructure, while DeepMind researchers will work with the studio to build out new workflows. The deal does not give Google access to A24's content library or its data. That last point is load-bearing: DeepMind is not acquiring training data. It is acquiring collaborative exposure to creative decision-making.

Google DeepMind invested $75 million in A24 to get inside the workflow that built it. Not the A24 library, but the A24 thinking. How A24 does it. The deal is a non-exclusive research partnership, which gives DeepMind access to A24's production process in exchange for development of AI infrastructure and tools. A24 partner Scott Belsky has indicated the tools will not resemble conventional prompted-generation AI, signaling an intent to build production-embedded capabilities rather than consumer-facing generative features.

It is seen as disappointing by some that a company benefiting from the anti-AI stance of its director Kane Parsons for "Backrooms" would make such a deal. A24 directors should "prepare to have your films altered against your wishes with this deal," one critic wrote. The fan backlash is notable because A24's brand equity rests partly on its reputation for filmmaker control. The deal represents the latest marriage between a Hollywood studio and AI in an era where companies have oscillated between partnerships and lawsuits.

The Appia Foundation is supported by a broad, cross-industry coalition including Arm, Armilla AI, Ericsson, Google, Mastercard, Microsoft, Mitsubishi Electric, Naaia, Nemko, Omron, OpenAI, Schneider Electric, and Siemens. The Linux Foundation announced the formation on June 17, 2026. Appia will develop open, modular specifications intended to translate international standards and established frameworks into practical assessment criteria across the AI value chain.

The project aims to create common specifications and assessment frameworks that organizations can use to demonstrate AI systems meet emerging safety, trust, and compliance requirements. The framework is designed to allow conformity evidence to be reused across the AI supply chain, potentially reducing duplicate assessments and compliance costs. Its work can help develop a critical missing trust layer by which third parties check conformity with standards, producing clearer and more reusable evidence when models, infrastructure, and applications are developed by different organizations.

The Appia Foundation's roadmap through 2026 includes the release of the full white paper and detailed specifications in August, the first expert-led webinar in September, and a conference or roundtable event in October and November. Critically, specifications are not yet published and Appia explicitly does not confer legal compliance, only conformity evidence for contractual and regulatory use. The Foundation produces conformity infrastructure; it does not confer compliance, and it does not claim regulatory authority. The membership spans model providers, deployers, assessment bodies, and insurers, giving Appia a broader supply-chain reach than prior safety forums.