AI Daily Brief — July 30, 2026

The day’s clearest signal is that AI is moving from experimentation into balance sheets and infrastructure commitments. Microsoft reported a sharp increase in paid Copilot adoption, Meta showed the cash cost of its AI buildout, and Europe opened a formal process to finance sovereign-scale compute; meanwhile, two new studies put useful limits around claims that today’s agents can autonomously conduct research or safely retain memory.

The essential updates

Microsoft says Microsoft 365 Copilot passed 30 million paid seats

What happened: Microsoft reported fiscal Q4 results on July 29 and said Microsoft 365 Copilot had exceeded 30 million paid seats, up from more than 20 million in the prior quarter. Azure and other cloud-services revenue grew 43% year over year, while total quarterly revenue rose 18% to $90.0 billion. Reuters separately reported quarterly capital spending above $41 billion, up more than 70% year over year, as Microsoft continued expanding AI capacity.

Why it matters: Paid-seat growth is stronger evidence of enterprise adoption than user registrations or demo traffic. It suggests Microsoft’s distribution through Office and Azure is converting AI demand into recurring revenue, while the Azure result shows that model training, inference, and related cloud workloads are materially affecting the broader platform.

What to keep in perspective: A paid seat does not show how frequently Copilot is used, whether it improves productivity, or whether customers will renew at the same scale. Microsoft’s cloud growth also includes non-AI workloads, and the capital intensity remains extraordinary: strong demand and strong unit economics are not the same claim.

Sources: Microsoft earnings release, July 29 · Reuters on cloud growth and capital spending · CNBC earnings coverage

Meta’s AI buildout is producing revenue growth—and consuming cash

What happened: Meta reported on July 29 that Q2 revenue grew 28% year over year to $60.8 billion, while net income fell 14%. Reuters reported that free cash flow dropped 91% to $784 million and that Meta narrowed its 2026 capital-expenditure forecast to $130 billion–$145 billion, compared with its earlier $125 billion–$145 billion range.

Why it matters: Meta is one of the largest buyers of AI infrastructure and a major open-model developer. Its results show both sides of the current cycle: AI-assisted advertising can support rapid revenue growth, but compute, data centers, talent, and long-duration commitments can absorb cash far faster than consumer AI products currently return it directly.

What to keep in perspective: Meta does not break out a clean “AI revenue” line, so it is not possible to attribute the quarter’s growth solely to AI. The company’s own earnings release was blocked by an automated security check in this run; the exact primary URL and figures were corroborated through the live earnings index and independent coverage.

Sources: Meta Q2 2026 earnings release · Reuters · CNBC

OpenAI shows that an agent benchmark can change dramatically with the harness

What happened: OpenAI published results on July 29 showing GPT-5.6 Sol’s ARC-AGI-3 public-set score rising to 38.3% when its Responses API retained reasoning between steps and compacted older context. OpenAI says the configuration used six times fewer output tokens than its earlier run. The official ARC harness had scored the model at 7.8%; independent coverage reports that ARC Prize co-founder François Chollet views general-purpose API settings as permissible but acknowledged a provider-parity problem.

Why it matters: The result is less about a new model than about evaluation design. Long-running agents can look dramatically weaker when a harness discards state they normally retain—and dramatically stronger when vendors tune execution around their own APIs. Buyers evaluating coding or operations agents should test the complete model-plus-harness system they will actually deploy.

What to keep in perspective: The 38.3% result is not directly comparable with every leaderboard entry because models were not necessarily run with equivalent context management, cost limits, or scaffolding. It is an OpenAI-run result on the public set, not an independent blind evaluation, and it should not be described simply as a new state of the art without those qualifications.

Sources: OpenAI technical note, July 29 · The Decoder’s report and ARC Prize response · ARC-AGI-3

Nscale is buying Anyscale to combine AI compute with the Ray software layer

What happened: Nscale announced on July 30 that it will acquire Anyscale, the company behind the managed platform built around the open-source Ray distributed-computing framework. Nscale says the deal will add orchestration and workload software to its data centers, GPU compute, and cloud services. Financial terms were not disclosed by the companies; Bloomberg reported a value of roughly $1.65 billion.

Why it matters: GPU access alone is becoming harder to differentiate. Owning the software that schedules training, inference, and agent workloads can improve utilization and make a provider more useful to developers. The transaction also puts stewardship and commercialization of a widely used open-source AI infrastructure layer inside a vertically integrated cloud company.

What to keep in perspective: Nscale’s “full-stack hyperscaler” language is positioning, not a demonstrated market outcome. Integration quality, customer portability, Ray’s open-source governance, and the economics of Nscale’s capital-heavy data-center expansion remain open questions; the reported purchase price is from unnamed-source reporting, not the announcement.

Sources: Nscale announcement, July 30 · Anyscale and Ray · Reuters · Bloomberg

The EU opens bidding for AI Gigafactories

What happened: The European Commission formally opened its AI Gigafactories call on July 30, seeking projects for very large AI compute sites and expecting the process to unlock about €30 billion in investment. Applications close November 12, with awards expected in early 2027; Reuters reports that the broader plan targets at least seven facilities and includes €10 billion in public funding.

Why it matters: Europe is moving from industrial-policy language to a procurement and financing process. If delivered, the facilities could give European labs and businesses more regional training and inference capacity, reduce dependence on US hyperscalers, and create demand for chips, power, cooling, networking, and data-center construction.

What to keep in perspective: A call for proposals is not built capacity. Site selection, permitting, grid connections, chip supply, private financing, and sustained utilization can all delay or weaken the plan. Sovereign compute also does not by itself create competitive models, software ecosystems, or profitable customers.

Sources: European Commission call, July 30 · Reuters · Wall Street Journal

Research worth noticing

Frontier agents completed the engineering but failed the research

A preprint submitted July 29 introduces “shadow evaluations”: agents receive the central question from an unpublished paper, then the original authors grade their work. In two case studies based on NeurIPS 2026 submissions, frontier agents received six days and thousands of dollars of compute. The authors report that the agents completed the engineering without human help but made no substantial progress on either research question; both outputs were rejected. Recurring failures included weak judgment about publishable novelty, poor backtracking, resource blindness, and instruction drift.

This is early evidence, not a population-level benchmark: two papers are far too few to establish a general capability ceiling, and the study has not been independently reproduced. Still, the setup tests a consequential claim more directly than short, automatically graded tasks. For non-specialists, the distinction is useful: automating experiments and code is not the same as choosing a good question, recognizing a dead end, or producing defensible new knowledge. The arXiv record links the paper; no separate public code repository was exposed on that record at cutoff.

Sources: Paper and authors

Persistent agent memory creates a delayed prompt-injection channel

MemSecBench, submitted July 29, tests 310 cases across 48 realistic code, science, daily-life, and office contexts using a controlled write–execute–forget protocol. Across 24 combinations of agent harnesses, memory backends, and models, the authors report that malicious content persisted in memory in 84.2% of cases and completed the full write-to-execution attack chain in 50.3%.

The benchmark’s contribution is lifecycle measurement: whether poisoned content is stored, recalled later, changes a real action, and can then be selectively removed. The figures are author-reported from a new preprint, use judge models at some checkpoints, and have not been independently reproduced; no standalone code repository was linked from the arXiv record at cutoff. The practical implication is straightforward: long-term agent memory should be treated as untrusted input with provenance, write controls, scoped retrieval, and deletion tests—not as harmless personalization.

Sources: Paper and benchmark description

Quick updates

• xAI announced Grok Voice Think Fast 2.0 on July 29 as its new speech-to-speech model; the official page was accessible through xAI’s news index but blocked direct browser inspection, so latency, pricing, languages, and independent quality remain unverified here. xAI
• Figma described an internal security agent that queries audit systems, investigates alerts, proposes code changes, and retains operational memory; Figma reports a 71% reduction in alert time-to-resolution, but this is a company case study rather than a controlled comparison. Figma, July 29
• AWS published a July 29 implementation guide for using Private Key JWT with Amazon Bedrock AgentCore Identity, relevant to teams that need machine-to-machine agent authentication without shared client secrets. AWS

The bottom line

• What changed today: Microsoft supplied a meaningful paid-adoption signal, Meta exposed more of the cash burden behind frontier-scale AI, and Europe moved its sovereign-compute plan into formal bidding.
• Who is most affected: Enterprise AI buyers, cloud and infrastructure operators, developers benchmarking long-running agents, and security teams giving agents persistent memory or production access.
• What deserves continued attention: Copilot usage and renewal—not just seat count; whether AI capital spending produces durable returns; fair model-plus-harness evaluation; and whether agent memory systems can prevent delayed, persistent prompt injection.