The OpenAI–AWS Deal Redraws the Cloud Map

---

Today’s AI cycle was all about infrastructure realignment. OpenAI effectively ending Azure’s exclusivity and going all-in on AWS is a move that shakes the balance of power in the enterprise AI market at its foundation. Gartner offered a bullish forecast of 47% growth in global AI spending this year, and AWS backed up the momentum with numbers of its own. Meanwhile, Figma used its conference to prescribe a remedy for “AI sameness,” and Japan’s Ministry of Defense published its approach to cognitive warfare. On the research side, a new framework for running reinforcement learning without ground-truth answers, and a paper arguing that world-model hallucinations are predictable and preventable, both stood out.

OpenAI × AWS: The “Azure-only” Era Is Over

On April 28, 2026 (US time), OpenAI and AWS announced an expansion of their strategic partnership. As a result, enterprises can now use OpenAI’s frontier model GPT-5.5, the coding agent Codex, and OpenAI-based managed agents directly on Amazon Bedrock. General availability for GPT-5.5 and GPT-5.4 began on June 1 (though, as of this writing, GPT-5.5 on Bedrock is limited to US regions).

Until now, OpenAI’s flagship models had been offered almost exclusively as a cloud service through Microsoft Azure. The turning point came the day before the announcement, on April 27, when Microsoft and OpenAI revised their contract: Azure priority is preserved, but it became clear that OpenAI could now “deliver all of its products through any cloud provider.” AWS moved immediately to capitalize on that opening.

The enterprise upside is clear. Companies can adopt OpenAI models while keeping their existing AWS security posture, IAM-based identity management, and procurement workflows intact. Tadao Nagasaki, head of OpenAI Japan, framed it this way: the heart of the partnership is letting customers “use the latest AI inside the familiar AWS ecosystem of security, governance, and observability.” The expected effect is to lower the “security-review wall” that often blocks AI adoption at large enterprises.

The growth rate of AWS’s AI platform, Amazon Bedrock, is itself a measure of the speed of the era. In his keynote at AWS Summit Japan 2026, Dave Brown revealed that the number of tokens Bedrock processed in the first quarter of 2026 alone exceeded the combined total of every prior period. AI agents, Brown said, are becoming “the largest workload on Earth.” Akihiko Shirahata, president of AWS Japan, also announced a plan to invest a cumulative ~¥3.8 trillion in domestic cloud infrastructure for the Japanese market by 2027.

Source: @IT - The “OpenAI is Azure-only” era ends: the upside of using GPT-5.5 and Codex on AWS ITmedia AI+ - Amazon Bedrock’s token volume in Q1 2026 alone surpassed its all-time cumulative total

Gartner Forecast: Global AI Spending Heads Toward $2.6 Trillion in 2026

A forecast Gartner published in May 2026 is drawing attention. Global AI spending in 2026 is projected to reach $2.6 trillion, up 47% year over year, and to expand further to roughly $3.5 trillion in 2027.

The single largest category is AI infrastructure, accounting for more than 45% of total spending — this covers AI-optimized IaaS, network fabric, and AI processing semiconductors. Spending on AI-optimized servers in particular is forecast to triple over the next five years. Gartner’s read is that “cloud providers are expanding capacity ahead of demand in anticipation of workloads from generative AI models and agentic workflows” — exactly the context behind AWS’s ¥3.8 trillion investment plan.

Spending on AI models is also surging. As the value of agentic automation becomes recognized, usage is widening, with a projected increase of $6 billion in 2026 alone — up 110% year over year. Gartner’s John-David Lovelock notes that “the potential value of agentic automation — executing multi-step processes and integrating with a wide range of tools — is now being recognized.”

Gartner sees 2026 as the inflection point at which mainstream enterprises that had been “waiting and watching” finally commit to serious investment. Companies that previously stalled at the proof-of-concept stage, unable to see the ROI, are changing their calculus as agentic automation becomes practical — and that shift is now showing up in the numbers.

Source: @IT - Is the corporate “wait and see” on AI spending finally over? Gartner’s forecast and the road ahead for serious investment

Figma Config 2026: How Companies Fight “AI Sameness”

When everyone uses AI tools, everyone makes similar things. Figma offered a prescription for this commoditization problem at Config 2026, its annual conference held in San Francisco on June 24–25.

According to Figma’s 2026 AI survey, more than three-quarters of people involved in product development say AI now lets them do work they couldn’t do before. Individual productivity is clearly rising. But as CEO Dylan Field put it, “as technology accelerates at an unprecedented pace, we now face existential questions about design and creativity.”

The real problem is team fragmentation. CDO Loredana Crisan says customers report that “AI has made individual work far easier, but collaboration has become completely impossible.” Because each member uses a different AI tool at their own pace, keeping a team aligned becomes hard — the side effect of AI is “higher productivity, fragmented teams.”

The core of Figma’s proposal: rather than handing the whole deliverable to AI, treat AI output as raw material that humans refine, and build mechanisms to share, across the whole organization, the AI techniques and house rules individuals develop — an approach that turns tacit know-how into an organizational asset. Figma says this can cut wasteful AI costs by up to 30%. Viewed as a startup opportunity, “tools that standardize and share an organization’s AI know-how” look set for rapidly expanding demand.

Source: ITmedia AI+ - How to overcome the “using AI makes you look like everyone else” problem: the keys to product differentiation

Japan’s Defense Ministry Publishes Its Approach to AI Fakes and Cognitive Warfare

On June 26, Japan’s Ministry of Defense released materials from discussions at its Defense Capability Transformation Task Force. The focus: countering “cognitive warfare” — using disinformation and psychological manipulation to influence an adversary’s perceptions and judgment.

The materials first analyze what happened during Russia’s invasion of Ukraine: the creation of imposter sites mimicking Western media, the spread of fabricated articles via fake social accounts, and the circulation of a deepfake video of President Zelensky calling for surrender. Ukraine quickly debunked the deepfake, but it still spread confusion and distrust for a time, the report notes. With generative AI now widespread, the cost of such attacks has fallen dramatically, and the Ministry explicitly warns that “advances in AI may accelerate the spread of disinformation, making cognitive-warfare response even more difficult.”

Ukraine’s counterattack is also instructive: establishing dedicated disinformation-monitoring organizations, running brand strategy in cooperation with ad agencies, fact-checking by private firms, and citizen-led OSINT (open-source intelligence) — combined to fight information operations. The structure of not just the military but also private companies and citizens working together to defend the information space offers useful lessons for Japan’s own response design.

Japan’s stated approach is to establish a “cognitive-warfare response cycle”: a three-phase loop of collect-and-analyze → disseminate → evaluate-and-improve, incorporating the latest technology including AI. The Ministry also says it will build R&D partnerships with the defense industry, universities, and startups in dual-use fields such as drones, cyber, and AI. The fact that the Defense Ministry is starting to view private AI startups as procurement targets is a market-opening opportunity worth watching.

Source: ITmedia NEWS - How will the Defense Ministry take on “cognitive warfare”? Its response to the AI fakes and fake accounts threatening Ukraine — analysis released

RiVER: A New Framework for Reinforcement-Learning LLMs Without Ground-Truth Data

Reinforcement Learning with Verifiable Rewards (RLVR) has become one of the main methods for training LLMs in recent years, but it carries a major constraint: it needs a ground-truth answer to assign rewards, so it can’t be applied to tasks where the answer isn’t uniquely determined.

A paper posted to arXiv — RiVER (Ranking-induced VERifiable framework) — attempts to break through that wall. Without any ground truth, it generates rewards by ranking the model’s own multiple candidate answers against each other, and uses that to drive learning. This opens the door to applying RLVR even to open-ended problems that have no single correct answer.

From a practical standpoint, the implications of being able to strengthen AI on “questions without a right answer” are significant. Legal document review, strategy support, creative feedback — none of these have a single correct answer. These domains have leaned heavily on human feedback (RLHF), which limits scale. If approaches like RiVER mature, that constraint could ease. That said, this is still an arXiv preprint, and practical-grade validation lies ahead.

Source: arXiv - Reinforcement Learning without Ground-Truth Solutions can Improve LLMs

World-Model Hallucinations Turn Out to Be “Predictable”

In world models — systems that let AI simulate future states, used in autonomous driving, robotics, and game AI — hallucination (simulation that diverges from reality) is a serious problem. Generated rollouts can look visually realistic while drifting away from actual physics and environmental dynamics.

A paper posted to arXiv, “Hallucination in World Models is Predictable and Preventable,” advances an important hypothesis: hallucinations don’t occur at random but concentrate in “low-coverage regions” of the state-action space — combinations of states and actions the model barely saw during training.

If that hypothesis holds, it becomes possible to predict and prevent hallucinations in advance. For example, you could surface a warning that “this simulation has entered a low-confidence area,” or collect additional data for that region and retrain. World models are increasingly used to make autonomous-driving experiments more efficient and to run physical simulations in medicine and manufacturing, so a practical method for controlling hallucinations should meaningfully accelerate commercial deployment. This too is a preprint, so replications and peer review are worth watching.

Source: arXiv - Hallucination in World Models is Predictable and Preventable

Wrap-up

The biggest story of the day is, without question, the OpenAI × AWS partnership. The axis of cloud competition is shifting, and enterprise AI adoption strategy is being freed from the “is it Azure or not” constraint. As Gartner’s $2.6 trillion forecast shows, this market is still in the early innings of its acceleration phase. Figma’s framing of the AI-sameness problem is an essential challenge the whole industry will face sooner or later. The contest of wits is moving from “what to build with AI” to “how to differentiate by orchestrating AI” — that, honestly, is the true face of today’s AI industry.