A North Korean threat actor (UNC1069) compromised the Axios npm library on March 31, slipping a Remote Access Trojan into OpenAI's macOS code-signing pipeline. Apps not updated by May 8 will stop receiving updates — OpenAI has found no evidence of user data compromise but urges immediate update.

Bloomberg and TechCrunch report Anthropic is fielding investor offers for a $50B round at a $900B valuation, which would leapfrog OpenAI's $852B post-money mark from March. A board decision is expected within weeks; the consideration is at an early stage with no commitment announced.

Released April 24, DeepSeek V4-Pro is now the largest open-weight model ever at 1.6 trillion parameters and a one-million-token context, benchmarking alongside GPT-5.4 and Claude Opus 4.6 at a fraction of the API cost. The training stack runs entirely on Huawei Ascend processors, bypassing U.S. export controls.

SoftBank is carving out a new U.S.-listed company called Roze to deploy autonomous robots for large-scale data center construction, targeting a $100B IPO as early as H2 2026. The entity absorbs SoftBank's acquisitions of ABB Robotics, Ampere Computing, and DigitalBridge.

Mistral AI launched Medium 3.5, a 128-billion-parameter dense open-weight model combining instruction-following, reasoning, and multimodal vision in one architecture. At $1.50 per million input tokens it significantly undercuts comparable frontier-tier competitors and is available on Le Chat and via API.

A Fortune analysis of Q1 earnings calls found Amazon, Google, Meta, and Microsoft have committed roughly $700B in AI capex for 2026 — more than the GDP of many mid-size nations — with no disclosed ceiling. Every hyperscaler cited demand outpacing available compute capacity.

CNBC reported a pending deal in which OpenAI would partner with Qualcomm to co-develop custom AI inference silicon for smartphones, sending shares up 7% in a single session. A successful alliance would give OpenAI a direct hardware path to on-device AI against Apple's Neural Engine and MediaTek.

In April alone, Maryland signed an AI dynamic-pricing bill, Tennessee enacted six AI measures, and Utah's governor signed nine AI bills covering deepfake protections, data provenance, and consumer transparency. The pace signals state legislatures are no longer waiting for a federal framework.

After the White House formally accused China of running "industrial-scale" campaigns to copy U.S. frontier AI via mass query distillation, a coalition of AI policy experts is calling for an immediate suspension of Nvidia H200 exports. No H200s have reached China yet, but the current case-by-case review posture leaves the door open — and critics say that's not enough.

Opus 4.7 posts meaningful gains on hard software engineering benchmarks and adds substantially higher-resolution image understanding. Simultaneously, Anthropic dropped the sales-call requirement for Enterprise plans, letting any organization self-serve at scale directly from the website.

Released April 22, Gemini 3 Deep Think is optimized for multi-step scientific reasoning and positioned as a direct competitor to OpenAI's o3-series on difficult STEM tasks. Available in preview via Gemini Advanced for subscribers.

Meta is cutting ~10% of its workforce and abandoning 6,000 open headcount slots, framing the move as an efficiency reset required to sustain its most aggressive AI capital cycle ever. The company projects $115–135B in capex for 2026 — roughly double what it spent in 2024.

CFO Sarah Friar confirmed a Q4 2026 debut timeline and said individual investors will receive IPO allocations — unusual at this scale. OpenAI's annualized revenue surpassed $20B by end-2025 and is projected to more than double in 2026, but the company still expects a $14B operating loss this year.

An April 24 OSTP memo accused foreign actors — primarily China — of systematically flooding U.S. frontier models with queries to extract training signal for competing systems. The Trump administration pledged tailored technology restrictions, though enforcement options remain limited without new legislative authority.

A federal court granted final approval to the Anthropic authors class-action settlement on April 23, after an earlier ruling that AI training on copyrighted books is fair use but storing pirated copies is not. The $1.5B deal sets an early precedent for how AI labs will price training-data licensing going forward.

Google launched two new Gemini Deep Research agent tiers on April 21, including Deep Research Max built on Gemini 3.1 Pro for extended multi-source research synthesis. The update makes multi-hour autonomous research workflows available to Gemini Advanced subscribers at no additional cost.

The Trump administration told Anthropic it will not allow the company to expand its restricted Mythos model to its planned 70-org consortium, halting the rollout of AI the UK safety institute confirmed can chain zero-days across every major OS and browser. The standoff pits Anthropic's voluntary disclosure framework against an administration asserting direct control over dangerous-AI deployment decisions.

GPT-5.5 and GPT-5.4 became available on Amazon Bedrock the day after OpenAI and Microsoft amended their partnership to drop cloud and IP exclusivity. Microsoft retains royalty-free access to OpenAI models through 2032; OpenAI is now free to sell on any cloud — Google Cloud and Oracle deals are expected to follow.

Alphabet's cloud division cleared $20B in quarterly revenue for the first time, powered by AI workloads, but management said demand outpaced available compute — suggesting underlying growth was even higher. The constraint underscores that data-center supply, not customer demand, is now the binding factor in the AI buildout.

Meta's Q1 2026 earnings call raised its full-year capex forecast to $125–145B — above prior guidance and analyst estimates — as the company accelerates AI infrastructure investment. Reality Labs continued to lose money, but AI-driven ad efficiency held revenue growth above 20% and management showed no sign of pulling back on the AI buildout.

Canadian enterprise AI firm Cohere and German sovereign-AI startup Aleph Alpha announced a merger, aiming to build an open-weights European powerhouse for regulated industries. The deal signals consolidation in the enterprise open-source tier as US frontier labs widen their capability lead and EU regulators push for sovereign AI alternatives.

On Microsoft's Q1 earnings call, CEO Satya Nadella said the restructured OpenAI partnership gives the company "the best of all worlds" — royalty-free access to all OpenAI IP through 2032 without exclusivity obligations. Microsoft's AI revenue hit a $37B annual run rate, up 123% YoY; Azure remains supply-constrained, not demand-constrained.

Rogo, the agentic AI platform built for investment banking, closed a $160M Series D led by Kleiner Perkins with Sequoia, Thrive, Khosla, and J.P. Morgan Growth Equity participating. The round is among the largest vertical-AI deals of the year and reflects Wall Street's push from AI pilots to production agentic deployments.

Reps. Ted Lieu and Jay Obernolte introduced the American Leadership in AI Act, folding over 20 bipartisan proposals into a single bill covering transparency, safety testing, export controls, and federal AI procurement. The omnibus approach is designed to break the legislative gridlock that has stalled standalone AI bills for three consecutive sessions.

Elon Musk was the first witness in his $130B breach-of-charitable-trust suit against OpenAI, testifying he "came up with the idea, the name, recruited the key people" and that Altman's nonprofit-to-for-profit conversion is "looting every charity in America." Altman, Brockman, and Microsoft CEO Satya Nadella are on the witness list this week.

The Wall Street Journal reported OpenAI fell short of internal revenue and weekly-active-user goals in early 2026, with CFO Sarah Friar warning compute commitments could become hard to fund. Oracle fell 4%, SoftBank sank 10%, AMD and Broadcom dropped 3–4% on the news. OpenAI called the story "clickbait" and said its business is "firing on all cylinders."

Beijing's state planner ordered Meta and AI-agent startup Manus to unwind their $2B acquisition following a months-long export-control review. Meta had already integrated Manus into internal systems and its executives had joined the company, making the unwind practically complex. The ruling signals China will use acquisition reviews as a strategic tech-war lever.

Following V4's April 24 launch, DeepSeek cut cache-hit input prices by 90% and added a 75% V4-Pro discount through May 5, bringing cost to $0.0036/M tokens vs. GPT-5.5's $0.50/M. The aggressive pricing targets US developer adoption and risks triggering a margin war across frontier AI providers. (New pricing development; V4 launch was covered April 28.)

Attackers compromised Context.ai, a third-party AI productivity tool used by a Vercel employee, then pivoted via OAuth to exfiltrate customer credentials. The stolen database appeared on BreachForums priced at $2M. The incident is one of six AI-adjacent supply chain attacks in the fifteen days since April 7.

European Parliament committees adopted amendments extending deadlines for stand-alone high-risk AI systems to December 2, 2027, and for AI embedded in regulated products (medical devices, vehicles) to August 2, 2028. The 18-month extension follows industry lobbying and reshapes enterprise compliance roadmaps across the bloc.

A paper at ICLR 2026 found that RL-based reasoning improvements increase tool-hallucination rates in lockstep with task gains — a fundamental tension in agentic AI design. Separately, Deloitte found 47% of enterprise users had made at least one major business decision based on hallucinated AI output.

Senators Warren and Blumenthal demanded Nvidia explain whether its $20B Groq licensing deal — which also transferred Groq's leadership team — was structured to evade Hart-Scott-Rodino antitrust review. Nvidia holds ~90% of the GPU market; Groq was among the few credible inference competitors. Senators urged the DOJ and FTC to open formal investigations.

DeepSeek released V4 Pro (1.6T params, 49B active) and V4 Flash on April 24 — one year after V3 blindsided Silicon Valley. Both support 1M-token context and open-source weights; V4 Flash input costs $0.14/M vs. GPT-5.5 Pro's $30/M. Benchmarks show near-frontier performance across coding, reasoning, and agentic tasks.

Alphabet is committing $10B now and up to $40B total in Anthropic — the largest single-investor AI commitment on record — plus 5 GW of compute via Google Cloud over five years. Anthropic's annualized revenue topped $30B, driven by Claude Code; with Amazon's parallel $20B commitment, Anthropic has both hyperscalers locked in as infrastructure backers.

Nvidia's stock rose 4.3% to $208.27 on April 24, pushing market cap above $5 trillion for the first time. The milestone reflects combined 2026 AI infrastructure spending guidance of $635–670B from Microsoft, Google, AWS, and Meta — direct GPU demand that has fueled a 14-fold share-price rise since 2022.

Intel posted its biggest single-day gain since 1987 after Q1 revenue of $13.58B beat estimates, with data center AI CPU revenue up 22% YoY. AMD added 14% on a DA Davidson upgrade. Analysts now project the semiconductor market crosses $1 trillion in 2026 — a third consecutive year of double-digit growth.

Novo Nordisk announced today a global partnership with OpenAI to integrate AI across all operations — from identifying obesity and diabetes drug candidates to manufacturing and commercial functions. Novo stock jumped ~3.5%; the deal signals enterprise AI moving from pilot to production across regulated industries.

The UK AI Safety Institute confirmed Anthropic's Mythos Preview can identify and chain previously unknown exploits across every major OS and browser — over 99% of discovered vulnerabilities remain unpatched. Anthropic gated access to internet-critical companies only and launched Project Glasswing for coordinated disclosure. (Core Mythos announcement April 7; AISI evaluation April 2026.)

PwC's 2026 AI Performance Study finds leading companies are capturing AI gains by reinvesting in growth, while the rest see flat returns. The widening gap is compounding faster than laggards can respond — a structural divide likely to reshape enterprise AI strategy discussions through year-end and into 2027.

Opening arguments in Musk's $134B breach-of-charitable-trust suit against OpenAI began today in Oakland federal court. The nine-juror verdict will be advisory; Judge Gonzalez Rogers decides remedies. The witness list includes Musk, Altman, Satya Nadella, and former OpenAI board members. (Update from April 27 coverage — trial now live.)

Released April 17, Grok 4.3 Beta adds native video processing and in-chat slide-deck generation. Initially gated to SuperGrok Heavy ($300/mo); wider rollout expected mid-May. xAI is simultaneously training Grok 5, targeting Q2 2026. (11 days old; included for model-landscape coverage.)

NVIDIA launched Ising, open-source AI models delivering quantum error-correction decoding 2.5x faster and 3x more accurate than classical approaches — bridging near-term quantum hardware and practical quantum advantage. An early signal that AI and quantum computing are converging faster than the field expected. (April 2026; exact date unconfirmed.)

China's National Development and Reform Commission today ordered Meta to unwind its ~$2B acquisition of Singapore-registered, China-rooted agent startup Manus — the most pointed blow yet to US-China AI deal flow. Meta has already integrated Manus into its agent stack, making the rollback messy.

OpenAI is now free to ship products on any cloud, Microsoft drops IP exclusivity (license stays through 2032 but non-exclusive), Microsoft no longer pays OpenAI a revenue share, and the AGI clause that gave Microsoft a contractual veto is gone. Six-year alliance formally loosens.

Meta signed the first commercial capacity reservation for space-based solar — up to 1 GW from Overview Energy's planned 1,000-satellite geosynchronous constellation that beams infrared down to Earth-based solar farms. Demo flight 2028, commercial 2030; aimed at AI data centers' 24/7 load.

Update on yesterday's coverage: Judge Yvonne Gonzalez Rogers seated nine jurors with no alternates today, and confirmed the jury verdict will be advisory only — she will decide remedies herself. Witness list now public: Musk, Altman, Satya Nadella, and current and former OpenAI board members. Opening arguments Tuesday.

Home-security firm ADT confirmed a data breach after ShinyHunters posted it to their leak site with an April 27 ransom deadline expiring today. Names, phone numbers, and addresses exposed; DOBs and partial SSN/Tax-ID for a smaller subset. Latest in a string of AI-augmented extortion campaigns.

Buried in Tesla's Q1 2026 10-Q is a single sentence describing a pending acquisition of an unnamed AI hardware company for up to $2B in stock and equity awards. Pairs with Musk's $25B 2026 capex commitment for Optimus, the AI5/AI6 silicon line, and Cybercab — Tesla's biggest known M&A move in years.

Hardware engineering chief John Ternus, 50, will succeed Tim Cook as Apple CEO on September 1; Cook moves to executive chairman after a 15-year run that took Apple from $300B to $4T. Ternus inherits Apple Intelligence's slow ramp and the company's persistent AI-talent gap.

Vercel disclosed a breach traced to a compromised Context AI vendor account; customer data is now being sold on BreachForums. The CEO publicly attributed the attackers' unusual velocity to AI augmentation — an early high-profile data point in 2026's AI-accelerated adversary discourse.

Yale seniors Nathaneo Johnson and Sean Hargrow raised $5.1M from Pear VC, Reddit CEO Steve Huffman, GPTZero founder Edward Tian, and Venmo's Iqram Magdon-Ismail to build a social network living entirely inside iMessage — an early test of whether AI-native consumer apps can ride iMessage's distribution moat.

CEO Evan Spiegel told staff in an April 15 memo that AI agents now generate over 65% of Snap's new code and field 1M+ queries a month, justifying a 16% workforce cut and $500M+ in projected annualized savings. Stock rose ~9% — the market is openly rewarding AI-driven payroll cuts. (Older story, kept for narrative weight; 7 of 10 items in this card are under 7 days old.)

Jury selection in Musk's $134B suit against Altman and OpenAI got underway today in federal court in Oakland; Musk dropped his fraud and constructive-fraud claims Friday, leaving only breach of charitable trust and unjust enrichment for the nine-juror, four-week trial. Opening statements begin no earlier than Tuesday.

GPT-5.5 ("Spud") rolled out April 23 to ChatGPT Plus/Pro/Business/Enterprise and Codex, with gains in agentic coding, computer use, and long-horizon tasks; it tops the Artificial Analysis Intelligence Index at 60 and roughly doubles the API price versus GPT-5.

Jeff Bezos's physical-AI lab (co-led with ex-Google scientist Vik Bajaj) closed a $10B round at $38B with BlackRock and JPMorgan; total funding now exceeds $16B. Bezos is reportedly seeking up to $100B more for a holding company to acquire industrial businesses and feed Prometheus their data.

Meta and AWS unveiled a multi-billion-dollar deal Sunday for Meta to run next-generation AI systems on Amazon's custom Graviton silicon, making Meta one of the largest external Graviton customers; Amazon shares rose ~3.5% on the news.

Cognition Labs, maker of the Devin AI software engineer, is in early talks to raise hundreds of millions at a $25B valuation — more than double September's $10.2B mark. Customers now include Goldman Sachs, Citi, Dell, Cisco, Palantir, and Mercado Libre.

xAI unveiled grok-voice-think-fast-1.0 on April 23, scoring 67.3% on the τ-voice Bench against Gemini 3.1 Flash Live (43.8%) and GPT Realtime 1.5 (35.3%). The model performs background reasoning so tool-call-heavy enterprise voice workflows don't pay a latency tax.

Google introduced TPU 8t and TPU 8i at Cloud Next on April 22 — first time the line has bifurcated. A 9,600-chip 8t superpod hits 121 exaflops; 8i triples on-chip SRAM and claims 80% better inference perf-per-dollar. GA later in 2026, direct Nvidia challenge.

Geoffrey Hinton told the UNRISD Digital World Conference that regulating AI without any "steering wheel" is more dangerous than driving without brakes. Yoshua Bengio's Scientific Panel report will feed the UN's Global Dialogue on AI Governance in Geneva in July.

At CadenceLIVE Silicon Valley on April 20, Cadence and NVIDIA expanded their partnership across agentic AI, multiphysics simulation, and digital twins. The combined stack pairs NVIDIA's Isaac/Cosmos world models with Cadence's high-fidelity physics for end-to-end physical-AI training.

AI-IDE maker Cursor is in talks to raise at least $2B at a $50B valuation led by Thrive and Andreessen Horowitz, with Nvidia and Battery joining; the round nearly doubles its $29.3B June 2025 mark and reflects revenue projected to exceed $6B by year-end.

Judge Yvonne Gonzalez Rogers approved Musk's motion Friday to streamline the case to two claims — unjust enrichment and breach of charitable trust — and jury selection begins Monday in Oakland for a four-week trial. Musk is asking the court to remove Altman and Brockman and unwind OpenAI's for-profit conversion.

DeepSeek shipped V4-Pro (1.6T params, 49B active) and V4-Flash (284B/13B) on April 24 under MIT license with a 1M-token default context. V4-Pro hits 80.6% on SWE-Bench Verified — within 0.2 points of Claude Opus 4.6 — at open-weights pricing that hammers frontier-lab API margins.

The Trump DOJ formally intervened Friday in xAI's federal challenge to Colorado SB24-205 — the first time DOJ has joined a case against a state AI regulation. DOJ argues the law's mandated bias assessments violate Equal Protection; SB24-205 is set to take effect June 30.

In a letter dated April 23 and surfaced this week, Sam Altman conceded OpenAI should have alerted RCMP about the future Tumbler Ridge shooter's flagged ChatGPT account in June 2025 — leadership had overruled employees who urged escalation. BC Premier David Eby called the apology "grossly insufficient."

Nvidia closed Friday at $208.27 (+4.3%) — its first record since October 2025 — pushing market cap past $5T. Intel had its best day since 1987 (+24%) and the SOX index notched its 18th straight gain, a broad-based vote of confidence in the AI capex cycle ahead of the May 20 earnings print.

Meta is laying off ~8,000 (10% of staff) starting May 20 to fund $115–$135B in 2026 AI capex. Microsoft, in its 51-year history's first voluntary buyout, opened the door to senior-director-and-below US employees with combined age + tenure ≥70. Big Tech is now openly converting payroll into compute.

Duke Energy unveiled a five-year $103B capex plan — an industry record — including ~20 new GE Vernova gas turbines and ~20 GW of new generation over a decade aimed at hyperscaler customers Amazon, Microsoft, Google, and Meta. CEO Harry Sideris said growth "is not just a blip."

OpenAI's standalone Sora app goes dark today (April 26); the Sora API follows on September 24. The discontinuation, announced March 24, ends OpenAI's social-video experiment and quietly signals compute is being pulled back to core enterprise products under cost pressure.

CSO Online reports CISA still has no access to Anthropic's bug-finding Mythos model — even after this week's third-party Discord-group leak — because Anthropic admitted federal cyber regulators to Project Glasswing only after AWS, Google, JPMorgan, Microsoft, Nvidia, Cisco, and CrowdStrike. Awkward private-first governance for a frontier offensive-cyber tool.

CNBC reports Chinese humanoid startups are already shipping units to factories and malls while US peers (Figure at $39B, Boston Dynamics/Hyundai, Tesla Optimus) sit on richer valuations and slower production. Honor's "Lightning" robot won the Beijing E-Town Half-Marathon on April 19 in roughly 50 minutes.

Alphabet committed $10B now in cash at a $350B valuation, plus $30B more if Anthropic hits performance targets, and Google Cloud will deliver 5 GW over five years. Lands days after Amazon's $25B add-on — Anthropic now sits between two hyperscalers it also competes with.

OSTP director Michael Kratsios issued NSTM-4 yesterday accusing China of using tens of thousands of proxy accounts and jailbreaks to distill capabilities from Claude, ChatGPT, and Gemini. Frames model distillation as a national-security threat and tees up new export controls.

A Discord group accessed Project Glasswing's preview Mythos — a cyber-offensive model Anthropic deemed too powerful to release — through a third-party vendor environment, partly using info from the Mercor breach. Continuous unauthorized access reportedly persists.

House Foreign Affairs voted out a slate of export-control bills aimed at closing loopholes that allow Nvidia H200 chips and frontier AI tools to reach China. Direct rebuke of the Trump administration's 25%-surcharge H200 deal — Congress is reasserting itself on chip policy.

Oklo's sodium-fast-reactor platform, NVIDIA's AI infrastructure, and LANL materials science combine to validate plutonium-bearing fuels for "AI factories." Oklo +15.65% on the news. Concrete signal that frontier compute and SMR policy are now joined at the hip.

SMR builder X-energy IPO'd yesterday at $1B, riding its 5 GW Amazon offtake deal and a Dow chemical-plant heat-and-power partnership. Second nuclear-AI vector this week — capital is voting that AI demand is real, durable, and power-bound.

Codex-powered, always-on agents now connect ChatGPT Business/Enterprise/Edu to Slack, Salesforce, Google Drive, Microsoft 365, and Notion. Free in research preview until May 6, then credit-metered — a direct play at the same enterprise seat Anthropic is winning with Claude Code.

Devin maker Cognition is in talks for "hundreds of millions or more" at a $25B valuation, more than 2× last September's $10.2B round. Post-Windsurf ARR has more than doubled — coding-agent multiples are holding firm even as Cursor closes its own $50B+ round.

Anthropic posted a Transaction Principal role focused on Frankfurt, London, Amsterdam, Paris, Dublin, the Nordics, and Southern Europe. With $100B+ already pledged to AWS and Google now offering 5 GW, this is Anthropic explicitly diversifying compute geography.

A federal securities suit and follow-on April 22–24 lead-plaintiff notices allege B&W's $2.4B "AI power" agreements were inflated and tainted by undisclosed BRC Group ties. Pushes AI-washing litigation past software and into the energy-infrastructure stack.

Amazon will invest $5B now and up to $20B more tied to milestones, on top of its prior $8B. Anthropic commits $100B in AWS spend over 10 years and will bring ~1 GW of Trainium2/Trainium3 online by year-end — a decisive bet on Amazon silicon over Nvidia.

At Cloud Next 2026, Thomas Kurian confirmed the Apple deal: a 1.2T-parameter custom Gemini on Apple's private servers, ~$1B/yr licensing. iOS 26.4's contextual Siri already uses it; the fully conversational version ships with iOS 27 in September.

NEC became Anthropic's first Japan-based global partner yesterday, rolling Claude Code and Claude Cowork to 30K employees and building industry-specific AI for finance, manufacturing, and local government. A direct challenge to OpenAI's and Google's Japan push.

Tesla beat Q1 at $22.38B revenue and raised its 2026 capex guide to $25B from $20B last quarter, explicitly to fund AI compute and Optimus factories. First-gen Optimus pilot at Fremont is targeting a 1M-unit annual run rate.

Anthropic filed for summary judgment in Concord v. Anthropic, arguing Claude's training on copyrighted lyrics is transformative fair use. Publishers' own motion is teed up; Judge Eumi Lee hears both July 15 in a case that will set the floor for music-AI licensing.

Cerebras filed its S-1 on April 17 targeting a mid-May IPO to raise more than $3B at roughly $35B. It disclosed a 750 MW, $20B+ OpenAI inference contract and an AWS deployment deal — a real dent in Nvidia's inference monopoly ahead of pricing.

Humble Robotics came out of stealth yesterday with a $24M Eclipse-led seed and a cabless electric freight hauler designed around 360° camera/LiDAR/radar for dock-to-dock operation. Another data point that physical-AI capital is shifting from robotaxis to freight.

Tesla told investors Optimus V3 debuts mid-2026 with mass production starting in Q3, using converted Model S/X lines in Fremont before moving to Giga Texas for a 10M-unit annual target. The aggressive ramp assumes perception and grasping issues are now solved.

Ex-Meta team Sooth Labs is raising a ~$50M Felicis-led round at a $335M valuation to build models that forecast geopolitical and market events. Unusual dual backing from Yann LeCun and Google's Jeff Dean signals "prediction models" as a new frontier-lab category.

On the Q1 call, Musk conceded that ~4M HW3 vehicles cannot be upgraded to unsupervised FSD and Tesla will offer discounted HW5 retrofits. A major admission that years of "all Teslas have full self-driving hardware" marketing was wrong — and a fresh liability surface.

DeepSeek dropped MIT-licensed V4-Pro (1.6T total / 49B active, 1M-token context) and V4-Flash on Hugging Face today. V4-Pro uses ~27% of V3.2's inference FLOPs and 10% of the KV cache while benchmarking near GPT-5.5 and Claude Opus 4.7 on math and coding.

Reports today say Musk's xAI has held recent talks with Paris-based Mistral and AI-coding startup Cursor about a joint push against OpenAI and Anthropic. Comes days after SpaceX's $60B option to acquire Cursor and signals xAI sees coding agents as the frontline.

Musk published a detailed Terafab plan today: $20–25B vertically integrated fab using Intel's 14A node, with SpaceX handling high-volume manufacturing. Prototype line at Giga Texas, full production "later this decade." A direct challenge to TSMC's AI-chip dominance.

New detail today on the DeepSeek funding story first reported April 22: Tencent has proposed up to a 20% stake in what would be DeepSeek's first outside round, but the lab is resisting giving up that much control. Valuation benchmark: MiniMax's ~$40B.

Jury selection starts April 27 on Musk's claim that Altman and Brockman defrauded him into $38M of donations by promising OpenAI would stay a nonprofit. Remedies sought include ousting Altman and Brockman and reverting OpenAI to nonprofit status. Four-week trial, Nadella to testify.

TechCrunch confirmed Delve — the YC-ousted compliance startup accused of fake audits — certified Context AI, the vector for the Vercel breach. A second Delve-certified customer was hit this week. The "AI compliance" layer is becoming a supply-chain liability.

New disclosure yesterday: customer data was exfiltrated from Vercel prior to the previously reported Context AI–linked breach. A Vercel database access key and env variables are reportedly listed for $2M on BreachForums. Materially widens the blast radius.

Fortune/Baker McKenzie count 51 AI-related securities class actions in five years, most alleging overstated AI capabilities. The suits are shifting from "does the AI exist" to "does it materially change unit economics" — a new disclosure standard is forming in real time.

SK Group announced today a partnership with Vietnam to build national AI infrastructure and industry-specific AI services — the first overseas deployment of its "Korean-style AI full-stack" model. A template for how mid-power economies plan to participate in the AI buildout.

Hugging Face released ml-intern this week: an open-source agent built on smolagents that autonomously runs literature review, dataset discovery, training scripts, and eval loops for LLM post-training. Reported to beat Claude Code on scientific-reasoning benchmarks.

Private-market buyers on Forge Global are now pricing Anthropic at roughly $1T — 2.6× February's $380B primary — after annualized revenue ran from $9B at end of 2025 to $30B by March. OpenAI trades at $880B on the same platform.

Amazon founder Jeff Bezos's physical-AI lab closed a $10B round with no lead, JPMorgan and BlackRock participating. Bezos is reportedly raising up to $100B to acquire industrial businesses and feed their operational data into Prometheus's models.

Google consolidated Vertex AI, Agent Studio, ADK, and Model Garden (200+ models) into a single enterprise agent stack pitched against OpenAI and Anthropic. New Gemini 3.1 Pro, 3.1 Flash Image, Lyria 3, and the A2A agent-to-agent protocol ship today.

Alongside the Agent Platform, Google pledged $750M to its 120,000-member partner network for joint customer transformations. Signals Google is buying distribution leverage rather than competing on model quality alone.

China's frontier-model lab is negotiating with Tencent and Alibaba after refusing VC money since inception; Tencent has proposed up to a 20% stake. A major shift for a lab that operated on quant-fund profits and open-sourced its frontier models.

Meta told employees today it will cut ~8,000 roles on May 20 and cancel 6,000 open reqs, reorganizing teams into AI-focused "pods" inside Applied AI. More H2 cuts expected as Meta pours $115–135B into AI infrastructure.

OX Security found a command-injection design flaw in Anthropic's official MCP SDK (Python, TypeScript, Java, Rust) — 200K+ servers, 7K+ public MCP instances, 14 CVEs already assigned. Anthropic's response, "expected behavior," is fueling industry debate.

Opus 4.7 now runs scheduled multi-step jobs on Anthropic's own web infrastructure — cron and GitHub-webhook-triggered agents with no CI host required. The shift turns Claude into a durable worker, not just a chat partner, inside the autonomous-agent race.

Max subscribers can now run a Mac as a 24/7 AI worker that touches the file system and drives native apps via a Command-Command shortcut. Perplexity recommends a dedicated Mac mini; $20 Pro users are excluded from the feature.

Intel posted Q1 revenue of $13.58B vs. $12.42B expected; Data Center and AI revenue hit $5.1B, up 22%. Shares spiked 20% after-hours as agentic workloads pull compute spend into CPUs alongside GPUs — a rare positive data point for "AI infrastructure ex-NVIDIA."

OpenAI released GPT-5.5 (codename Spud) just six weeks after GPT-5.4. New state of the art on Terminal-Bench 2.0 at 82.7%. API pricing doubled to $5 / $30 per million in/out tokens.

SpaceX offered Cursor either a $60B buyout later this year or a $10B collaboration payment, derailing the AI coding startup's Andreessen-led $2B funding round at a $50B valuation.

At Cloud Next 2026, Google announced two new AI chips. Training performance 2.8x prior Ironwood TPU, inference up 80%. Supports single-cluster scale past 1 million TPUs.

Anthropic identified a reduced reasoning default, a caching bug, and verbosity prompt changes behind weeks of Claude Code quality complaints. Usage limits reset for all paid subscribers today.

Sony AI's Ace project, published in Nature today, is the first autonomous system competitive with expert humans in a physical sport. A landmark in real-world robotic perception and control.

OpenAI's first image model with native reasoning claimed #1 across every Image Arena category within 12 hours of launch. DALL-E 2 and 3 are being retired May 12, 2026.

Grok has thrown "high demand" errors across the US, UK, and Germany since April 21. Paid SuperGrok subscribers and Ani Companion users are locked out alongside free users.

Attackers compromised a Vercel employee's Google Workspace through Context.ai, then pivoted into Vercel's environment to enumerate and decrypt non-sensitive env variables. Classic AI supply-chain attack.

Capital is flowing away from standalone AI software toward the power infrastructure AI needs. Blue Energy, EVAS ($211M RISC-V AI chips), and Syenta ($36M chip packaging) all closed big rounds April 22.

New report confirms what the Vercel incident showed in real time: unsanctioned AI tools and AI-targeted supply-chain attacks are rewriting the bank cybersecurity threat playbook for 2026.

New state of the art on Terminal-Bench 2.0 and OSWorld-Verified. API pricing doubled. A coding gap to Claude Opus 4.7 that OpenAI won't put on its slides.

75% on OSWorld vs. the 72.4% human baseline. The first AI to outperform humans at using a computer natively — not a plugin, not a preview. Native.