AI is giving the CIO a bigger mandate than ever before.
For decades, IT meant keeping the lights on: uptime, networks, devices, enterprise apps, and SLA compliance. The asset was infrastructure, and the function was judged on cost and ticket volume. In the AI era, the assets become data, models, and workflows, and the job moves from “we manage systems” to “we enable intelligence across the enterprise.”
The future CIO runs that system. They sit alongside the CTO, the CDO, and the CFO as a business co-creator and transformation leader, and the defining question changes from “are systems stable and secure?” to “is the company becoming AI-native, safely and fast enough?”
The CIO’s five roles in the AI era:
Enterprise AI Orchestrator – Decide which AI models are used and where
Owner of the AI Stack – Own the AI stack across data, models, applications, security, and governance
Enabler with Guardrails – Enable AI adoption with approved tools, policies, and platforms
Risk Owner, the CISO dimension – Govern AI security, risk, reliability, and compliance
Architect of the Human + AI Workforce – Design how humans and AI work together across the enterprise
Agents take over execution, but the advantage still belongs to the leader who sets direction, defines standards, and governs trust, security, and risk across all models and agents.
The CIOs who come out of this strongest will build the fastest, safest way for their company to use AI: a platform company within the company, where a small, high-caliber team enables everyone else to build safely. Their orgs will be flatter and more leveraged, their governance will be strong without being slow, and the function will shift from managing systems to enabling intelligence across the enterprise.
The question for every CIO: Are you becoming the fastest, safest way for your company to use AI, or the gatekeeper it routes around?
Below, we round up the signals that shaped the week: AT&T running open models for a quarter of its AI usage at 45 billion tokens a day, IBM putting OpenAI’s models inside its global consulting business and training tens of thousands of consultants to deploy them, Mistral hosting third-party open models under the same regional controls it applies to its own, and Anthropic adding machine-readable watermarks to Claude output as EU rules reshape products well beyond Europe.
Full Weekend Edition below. 👇
Signals Shaping the Future of AI:
Infrastructure
Microsoft targets September unveiling of next-generation Maia 300 AI chip. Sources say Microsoft plans to unveil its next-generation AI chip, the Maia 300, as soon as September, and is in talks with TSMC to produce more than 300,000 units for 2027. Click here
Google develops “Frozen” chip to cut Gemini inference costs. Google is building a server chip, internally called Frozen v2, that hardcodes the Gemini model architecture directly into silicon to cut AI inference power use by six to ten times per token. Click here
IBM and Together AI sign a $240 million AI inference cluster deal. The multiyear agreement builds an AI inference cluster on IBM Cloud using NVIDIA’s HGX B300 systems to support open source model deployment. Click here
Oracle’s AI data centers face multibillion-dollar cost overruns. Oracle’s AI megacampuses in Wisconsin and El Paso, including its Project Jupiter site in New Mexico, are running billions of dollars over budget as permitting disputes and a forced switch from natural gas to fuel cells raise costs. Click here
Anthropic signs a $9.1 billion, 20-year compute deal with Riot Platforms. Anthropic will pay for 191 megawatts of data center capacity at Riot’s Rockdale, Texas, campus over two decades; Riot Platforms shares jumped more than 15 percent on the news. Click here
SK Hynix commits to a $720 billion memory buildout at Yongin, South Korea. The cluster is set to become the world’s largest network of memory factories, entering production in February 2027, as SK Hynix races to meet AI-driven demand for high-bandwidth memory chips. Click here
Enterprise
IBM partnered with OpenAI to expand enterprise deployment of GPT-5.6, Codex, and ChatGPT Work through its global consulting business. IBM will train tens of thousands of consultants and build industry-specific AI solutions across financial services, government, telecommunications, and retail. Click here
AT&T says open-weight models now power a quarter of its AI usage. The company expects open models to account for 70 to 80 percent of its total AI usage over time, with an average of 45 billion AI tokens per day. Click here
Microsoft begins merging its consumer and commercial Copilot apps into one. The unified app rolls out on mobile and web in mid-August and on desktop in mid-September, ahead of a broader “super app” push. Click here
Google launched Gemini 3.7 Flash, a lower-cost model designed for coding and agentic business workflows. The model improves debugging, issue resolution, and production-ready code generation while launching at half the introductory cost of Gemini 3.6 Flash. Click here
Meta releases Muse Glimmer, a 30 billion-parameter open-weight model that runs on one consumer GPU. The Apache 2.0 model is distilled from Muse Spark 1.2, compressed to about 17 gigabytes for local agent tasks like scheduling and coding; Meta plans to open-weight Spark 1.2 too. Click here
Mistral will run third-party open models on its own infrastructure under the same regional controls as its models. The company starts with Z.ai‘s GLM-5.2, offering shared compliance and service commitments to enterprise customers. Click here
Capital Flows
Databricks closed a $5 billion round at a $190 billion valuation. The round comes six months after Databricks raised $5 billion at a $134 billion valuation, and the company says it has crossed $7 billion in annualized revenue run rate. Click here
Accel raises $3.5 billion to back early-stage AI startups globally. The fund includes a $1.35 billion global expansion vehicle for larger early-stage rounds and rapid follow-on investments in emerging AI companies. Click here
Thrive Holdings raised $2 billion at a $12 billion valuation to buy traditional businesses and rewire them with AI. SoftBank, D1 Capital Partners, and Altimeter Capital backed the round, bringing the OpenAI-linked roll-up’s total funding to more than $3 billion. Click here
AI code review startup CodeRabbit raised $143 million at a $1.5 billion valuation. The Series C was co-led by Atomico and Smash, following a $60 million Series B just eleven months earlier. Click here
River AI raised $1.1 billion just two months after launch to build personally trainable AI agents and a new end-to-end AI stack. The company is targeting enterprises that want more control over open models, with reinforcement learning and fine-tuning tools designed to reduce dependence on closed-source systems. Click here
Research
An AI agent powered by Claude Opus 4.6 exploited a gym reservation system to move its user up a waitlist without being explicitly told to hack it. The incident shows that even older agentic models can autonomously discover and exploit real-world vulnerabilities while pursuing routine user goals. Click here
NVIDIA is developing Nemotron 4, an open model exceeding 1 trillion parameters. The model would surpass Nemotron 3 Ultra’s 550 billion parameters, though it remains smaller than leading Chinese open-weight models, according to sources. Click here
DeepSeek launches V4-Pro, a flagship model priced well below rival Kimi K3. The model rivals Kimi K3 on some benchmarks while costing far less, at $0.435 per million input tokens and $0.87 per million output tokens. Click here
Anthropic says an unreleased Claude model made unexpected progress toward the Riemann hypothesis. The model didn’t solve the century-old problem but advanced a related question, Anthropic said in a research write-up. Click here
Policy
Anthropic is adding machine-readable watermarks to Claude-generated text and provenance metadata to AI-generated images. The move shows EU transparency rules beginning to reshape frontier-model products beyond Europe itself. Click here
Local bans on new data centers have surged past 500 US jurisdictions. New York and Texas joined the pushback with statewide restrictions, up from roughly 300 bans in late June, as communities cite power and water strain from AI buildouts. Click here
NIST opened public comment on a new framework for testing AI systems. The TEVV-Athlon Framework gives agencies and companies a standardized method for evaluating language, multimodal, and agentic models before deployment. Click here
A federal appeals court cleared roughly 2,400 lawsuits against Meta, Google, TikTok, and Snap to proceed. The Ninth Circuit rejected the companies’ bid to use Section 230 to block claims that they designed algorithmic feeds to addict young users. Click here
Trump signed a memo authorizing private firms to conduct offensive cyberattacks abroad. Vetted companies can now surveil and disrupt foreign criminal groups under federal oversight, provided they maintain at least a $1 million compliance bond. Click here
Global AI Strategy
Apple has trained a China-specific AI model with Alibaba’s support as it prepares to launch Apple Intelligence in the country. The model gives Apple greater control over its AI experience in China while navigating local regulatory requirements and competition from domestic smartphone makers. Click here
South Korea launches a $3.5 billion fund to build out its chip supply chain. The fund backs semiconductor materials, parts, equipment, and fabless companies, with another $3.5 billion in trade finance for suppliers, as President Lee pushes a chip megaproject exceeding $576 billion in planned investment. Click here
Microsoft has shuttered 15-plus branch offices and joint ventures in China over five years. The retreat comes as Beijing pushes domestic software adoption, though Microsoft says AI demand from Chinese firms expanding abroad keeps a window open. Click here
Foreign investment in Japan’s chip industry has reached $37 billion. Sony and TSMC’s planned image-sensor joint venture pushed the total higher, with Japan’s government considering financial support as it competes for AI-driven chip manufacturing. Click here
Taiwan says it detected an AI-assisted cyberattack campaign against government agencies in July. Officials traced the intrusion to an “overseas source” and say the targeted agencies successfully contained the attacks. Click here
Talent Signals
Each week, we spotlight key roles tied to the themes shaping this week’s AI headlines, connecting talent to the companies driving the news.
BigPanda is building an agentic AI platform for IT operations, helping enterprise teams detect, investigate, and resolve incidents while automating complex operational workflows. BigPanda is hiring across engineering, product, operations, and go-to-market roles. Open roles are listed on its careers page. Click here
PrimeIntellect is building an open superintelligence stack that gives AI teams access to training infrastructure, reinforcement learning environments, evaluations, and distributed compute traditionally available only inside frontier labs. Fresh off a $130 million Series A, the company is hiring across training infrastructure, research, compute, and technical customer roles. Click here
RiverAI is rebuilding the stack for personally owned AI, spanning model training, reinforcement learning, local inference, custom hardware, and consumer product experiences. Just two months after launching, River raised $1.1 billion and is hiring across frontier research, software engineering, product, data, and custom silicon. Click here
You can see all the opportunities at Mayfield-backed AI companies here.
Social Signals
The most important conversations in AI are unfolding across social media, where top voices are shaping the next wave of signals and strategy. Here are some of the top social signals and their takes from the past week.
Jensen Huang (Click here) — “A new asset class is being born. AI factories are becoming investable infrastructure. The capital markets are mobilizing to build the infrastructure of intelligence.” Huang announced NVIDIA partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR designed to mobilize more than $500 billion of third-party capital for AI infrastructure. His argument is that compute is evolving from a technology expense into a financeable, revenue-producing infrastructure asset, opening a new layer of opportunity across AI factories, power, networking, and the broader physical AI stack.
Alexander Panfilov (Click here) — “We found a way to extract the hidden reasoning of frontier models using a vulnerability in the APIs of every frontier AI company. We verified that our reasoning token count matches the billed API thinking tokens 1:1 for most of the prompts we queried.” In a post viewed 3.3M times, Panfilov shared research showing that concealed reasoning traces from proprietary models could potentially be recovered through API behavior. The work raises a new security question for frontier labs: protecting model outputs may not be enough if the underlying reasoning itself can leak through the infrastructure serving them.
Andrew Ng (Click here) — “Using agentic coding effectively is now a key skill for every developer. Given a clear spec, coding agents are rapidly improving at delivering to it. Thus, our work as engineers is shifting toward deciding what should be in the spec.” Ng introduces an AI Engineering Skills Map built from more than 10,000 job postings and interviews with AI experts and hiring leaders. His takeaway is that as agents handle more implementation, the highest-value engineering skills are shifting toward evals, software fundamentals, product judgment, and knowing what to build.
