Source: Gemini

The potential of AI has enterprises and hyperscalers alike racing to capture value — but the real unlock isn’t the technology. It’s solving privacy and security. Here’s the storyline in a few charts.

Hyperscalers are placing a $650B bet on AI infrastructure — and they need enterprise demand to justify it. The investment in AI compute and infrastructure is historic. The return depends on enterprises moving from experimentation to production at scale. link

Source: Ben Evans

But there is a significant gap between that massive infrastructure spend and actual enterprise AI spending, which sits at just $37B in 2025. The infrastructure is being built. Enterprise adoption is growing fast, but it pales in comparison to the infrastructure investment. Closing that gap is the central challenge for the AI ecosystem. link

Source: Menlo Ventures

Companies are seeing real ROI from AI — but it’s concentrated in a narrow set of use cases. The value is there. The opportunity is to expand it across a wider range of high-impact business problems. link

Source: a16z

The key to unlocking broader ROI is automation — not just AI assistance, but redesigning how work gets done. High-performing AI implementers are much more likely to be redesigning workflows, including automation. link

Source: McKinsey

Smart automation requires agents — and enterprise adoption is already accelerating rapidly. AI agents operate with increasing autonomy, executing multi-step tasks across systems. The question is no longer whether enterprises will use agents, but how fast. link

Source: G2

However, agents introduce a new category of risk that most organizations are not prepared for. AI tools are deployed in 73% of organizations, but real-time governance covers only 7%. Only 9% of organizations can intervene before an agent completes an action. link

Source: Cybersecurity Insiders

The result: data security and privacy is now the #1 concern for executives — and one of the few concerns that is still growing in importance. This is not a theoretical risk. It is the primary barrier between where enterprises are today and where they need to go. link

Source: Bain & Company

The companies that solve this become the ones that win. PwC finds that 58% of executives report improved AI ROI from responsible AI programs, and organizations with mature governance are significantly more likely to rate their AI capabilities as effective. link

Source: PwC

Privacy and security aren’t the brakes on AI progress. They’re the accelerant. A comprehensive governance approach integrated with run-time security is what turns AI ambition into durable enterprise value.

Source: Gemini

Hyperscalers make big bets on supply — and big waves driving demand

According to Bridgewater, Big Tech hyperscalers plan a massive $650 billion AI infrastructure investment in 2026. Yet Menlo Ventures reports 2025 enterprise AI spending hit just $37 billion. That gap suggests suppliers expect a major surge in AI service spending. Or it’s just FOMO. Or both. Reuters Menlo Ventures

Source: Ben Evans

To spur demand, Microsoft announced Copilot Cowork — an agentic application built with Anthropic’s help to automate non-engineering workflows. The agent executes plans in the background using Microsoft 365 data from Outlook, Teams, and SharePoint, while operating within enterprise security, privacy, and governance boundaries. link

The demand equation extends beyond direct enterprise implementations. Alphabet reported Google Search revenue grew 17% year-over-year to $63 billion, driven by AI Mode queries that have doubled since launch and run three times longer than conventional searches — creating new ad inventory as AI reshapes how people seek information. link

Security moves to center stage

That wave of new hyperscaler AI tools has already rattled markets. Announcements of Claude Code Security, Google CodeMender, and OpenAI Aardvark triggered a cybersecurity stock selloff — what Forrester called a “SaaS-pocalypse.” The anxiety isn’t unfounded: Google reports only 27% of security leaders feel equipped to secure applications against autonomous, agent-driven coding. Forrester report

Source: Forrester Research

The OpenClaw craze illustrates the stakes. The unleashed local agent sold out Mac Minis and racked up massive downloads — but also wiped a security researcher’s inbox against her explicit instructions. OpenAI has since hired founder Peter Steinberger to build next-gen personal agents, while transitioning OpenClaw to a foundation-supported open-source project. X Cisco

Source: Summer Yue/Simon Willison

The tide seems to be rising

The ROI picture remains murky. An Andreessen Horowitz survey finds enterprise AI returns mostly hovering between 1x (breakeven) and 1.5x — solid if not spectacular. link

Source: Andreessen Horowitz

An NBER study of 6,000 global executives confirms the pattern: little AI impact to date, though firms predict three-year gains in productivity (1.4%) and output (0.8%), paired with a 0.7% employment drop. Employees are more optimistic, expecting a 0.5% employment boost. Notably, CEOs use AI more than the rest of the C-suite. link

Source: NBER

To drive results, Google developed a KPI framework for production AI agents spanning operational reliability, adoption, and business value. Gartner adds five board-ready metrics — sales conversion rates, labor costs, time to value, collection efficiency, and employee NPS — to make the ROI case stick. Google Gartner

Wrestling with productivity

Harvard Business Review reports that AI is intensifying work, not reducing it. Workers using AI tackle a broader range of tasks with less downtime and more multitasking. Getting this balance right is critical for both employees and the employers expecting efficiency gains. link

Agency economic models are under pressure to evolve. S4 Capital’s Monks expects 25% of its revenue to be subscription-based by year-end — the logic being that as AI improves productivity, clients receive more output for the same fee rather than fewer billable hours. link

Working on workforce

Bain argues unlocking AI value requires simultaneously redesigning workflows and modernizing workforces. As agents deploy, work shifts from doing to planning and reviewing. The firm also notes that AI failures are often design flaws, not model errors — and perceived AI disruption risk is rising across all industries. link link

Source: Bain

Block is cutting 40% of its workforce and attributing it to AI. Wary observers suspect it’s mostly right-sizing after the company doubled its headcount between 2020 and 2024. Both can be true. link

Swings and roundabouts

Enterprise vendor trust is being tested. Anthropic refused Pentagon demands for unrestricted use of Claude and was blacklisted by the Trump administration. OpenAI signed a similar deal the next day, triggering 1.5 million #QuitGPT cancellations in 48 hours. For enterprise buyers, this raises real questions about vendor alignment. Meanwhile, Anthropic hit a $19B annualized run rate and Claude jumped to #1 on the App Store — the market rendering its own verdict. Anthropic CNBC CNBC Euronews Yahoo Finance

OpenAI’s agentic commerce ambitions have hit their own friction. After launching in-app shopping to much fanfare, the company pulled back its “Instant Checkout” feature — scaling down to referral links rather than native transactions. According to Forrester, direct buying within answer engines remains consumers’ least-adopted AI use case. link

I recently helped a client push two AI initiatives into live testing. The technology worked. What almost didn’t was the coordination — and the pattern is instructive because it applies broadly.

Getting an AI agent from prototype to production requires four parties in tight lockstep: domain experts making decisions on critical functionality, IT enabling data access and services, AI developers tuning agent accuracy and usability, and program management keeping everything synchronized. Remove any one from the cadence and progress fragments — not from a technical failure, but a coordination one.

What kept these two initiatives moving:

1. Weekly status focused on achievement, not problem-solving. Every party in the room, every week. Keep it tight; surface blockers and route them, don’t analyze them.
2. Active interim follow-up. Four interdependent teams can’t afford to wait until next week. Gaps compound fast.
3. Disciplined scope. Generative AI invites endless “what if” conversations. Aspirations go in the backlog. Phase deliverables stay fixed.
4. Time-bound accountability. Every deliverable has an owner and a date. No ambiguity about who owes what by when.

None of this is new. It is the same discipline that gets any cross-functional initiative across the line. But AI amplifies the coordination cost because the work spans domains that rarely share a natural operating rhythm — domain experts, IT, and AI developers operate on fundamentally different cadences and modes. That gap doesn’t close on its own.

The organizations shipping AI into production aren’t necessarily the ones with the best models. They’re the ones that treat orchestration as a core competency.

Source: Bain

The broader shift is accelerating: 74% of firms now rank AI a top-three strategic priority, up from 60% a year prior. Software development leads production scaling at 40%, customer service at 32%. In my client work, I see the same compounding dynamic — capabilities built for one function, say knowledge retrieval or invoice processing, transfer to others at lower cost and risk. Scale breeds scale.

Satisfaction rises sharply as companies move from AI-as-assistant to agentic workflow automation. Bain found respondents using agentic approaches were twice as likely to report results exceeding goals. But the risk profile shifts with it: security and privacy concerns climbed from 38% to 45% year-over-year, spiking to 49% among companies in production versus 35% still piloting. Automation unlocks the upside — it also demands tighter governance.

When use cases disappoint, the model is rarely the problem. Bain found a third of failures scaled poorly because designs didn’t accommodate real-world data and workflow complexity. Capturing those intricacies upfront — through shadow deployment and gradual rollout — separates durable production from a stalled pilot.

Leading firms share a common posture: defined vision, prioritized use case list, dedicated investment, and an iterative roadmap. Bain found 55% of leaders have all four in place, versus 40% of laggards who lack even a defined vision. The discipline isn’t just planning — it’s maintaining focus on shipping current initiatives while systematically identifying the next wave. That compounding effect is where the real ROI emerges.

Bain: Executive Survey — AI Moves from Pilots to Production

Source: Ken Calhoon

Now consider what happens when AI moves into knowledge retrieval, invoice matching, or warranty validation. There is no compiler. No automatic pass/fail. The default iteration cycle becomes: developers tweak the system, sample outputs get generated, domain experts review them days later, and the cycle restarts. That loop is brutally slow — and the further AI scales beyond IT, the more consequential this drag becomes.

The fix is a golden evaluation set built before development begins: 50 to 100 examples covering standard flows and realistic edge cases, with expected inputs and desired outputs defined for substance, format, language, and latency. This creates a synthetic feedback loop that replicates the speed advantage coding environments enjoy natively.

The quality benefit is obvious. The speed benefit is routinely overlooked. With automated evaluation in place, developers receive immediate pass/fail signals on every iteration. Refinement cycles compress from days to minutes.

The catch is real: domain experts must invest meaningful time before a single line of code is written. That feels counterintuitive under pressure to start building. But hours of upfront definition save weeks of slow, reactive review cycles. The math is unambiguous.

Leaders should make this a hard prerequisite: no use case enters development without a golden test set and a documented workflow or SOP. The upfront discipline is not a brake on momentum — it is what makes downstream velocity possible.

Uncovering insights requires structure, not serendipity. Here is the framework I use to filter the AI ecosystem:

• Curate Aggressively: Limit inputs to high-authority voices.
• Leverage Audio: Podcasts offer off-the-cuff insights you won’t find in written copy.
• Automate Everything: I run GitHub Actions to fetch, parse, and email me summaries daily.

My top five go-to sources span business strategy, news digests, technical research, and practical application:

The full curated list is below, organized by topic. Sources marked Featured are my highest-value picks.

AI News Digests & Weekly Roundups

Featured — Benedict Evans

Newsletter
Weekly curation and analysis of what happened in tech that mattered.
Author: Benedict Evans
Subscription: Freemium (free Tuesday edition; $150/yr for premium Sunday edition)

If I read one source per week to keep track of trends, this is it. A collection of major happenings, data, etc.

The Batch

Newsletter
Weekly digest of the most important AI research, industry news, and practical insights.
Author: Andrew Ng / DeepLearning.AI
Subscription: Free

Useful weekly digest of major trends across AI.

The Deep View

Newsletter
Tri-weekly summary of the most important AI news, tools, and tips in under 5 minutes.
Author: The Deep View
Subscription: Free

State of AI

Newsletter
Monthly synthesis of key developments in AI, AI markets, and geopolitics plus the annual State of AI Report.
Author: Nathan Benaich
Subscription: Free

Investor’s monthly perspective across disciplines, global.

AI Strategy & Industry Conversations

Featured — Beyond The Prompt

Podcast
How organizations leverage AI to accelerate success, with creative strategies and actionable tactics.
Presenters: Jeremy Utley & Henrik Werdelin

Long form, very interesting and wide ranging guests and discussions. Bigger thinking and focus on enterprises, but also some practical tips.

No Priors

Podcast
Leading AI engineers, researchers, and founders discuss the biggest questions in AI and technology.
Presenters: Sarah Guo & Elad Gil

Sharp Tech

Podcast
How technology works and the ways it is impacting the world, with weekly discussions.
Presenters: Andrew Sharp & Ben Thompson

Tech Unheard

Podcast
One-on-one conversations with industry leaders exploring pressing technology trends from AI to leadership.
Presenter: Rene Haas (CEO of Arm)

Interesting interviews with tech leaders; not necessarily about AI.

Business Strategy & Enterprise AI

Featured — Stratechery

Newsletter
Weekly and daily analysis of the business, strategy, and impact of technology.
Author: Ben Thompson
Subscription: Freemium (free weekly; $12/mo or $120/yr for daily updates and podcasts)

Highly condensed, covering major tech industry trends including AI.

Featured — One Useful Thing

Newsletter
Practical implications of AI for work, education, and everyday life from a Wharton professor.
Author: Ethan Mollick
Subscription: Free

Wharton professor’s incredibly popular blog; practical perspectives.

Featured — Applied AI

Newsletter
Reporting on how businesses are deploying AI to automate work across industries.
Author: The Information
Subscription: Paid ($199–399/yr subscription to The Information)

The Information is expensive, but this newsletter is free. Practical business application and implications.

Featured — Bain Software

Newsletter
Technology and software industry research and insights from Bain’s consulting practice.
Author: Bain & Company
Subscription: Free

They say in an email what others write a whole report on.

Enterprise AI Executive

Newsletter
Twice-weekly executive-level insights on generative AI, agentic AI, and enterprise AI strategy.
Author: Enterprise AI Executive
Subscription: Free

‘Executive’ focused, leans on research from major consultants and SIs.

The Change Constant

Newsletter
Weekly teardown of the biggest news in tech and finance for professionals.
Author: Saanya Ojha
Subscription: Free

Somewhat snarky business POV on AI.

Me, Myself, and AI

Podcast
MIT Sloan Management Review and BCG collaboration sharing practical AI implementation secrets and success stories.
Presenters: Sam Ransbotham & Shervin Khodabandeh

The So What from BCG

Podcast
BCG leaders on trends, developments, and ideas that will shape and disrupt the future of business.
Presenter: Georgie Frost

Practical AI Application & Tips

Featured — How I AI

Podcast
Live screen-sharing episodes where guests demonstrate specific, practical AI workflows you can copy immediately.
Presenter: Claire Vo

Very practical on usage of AI by bleeding edge practitioners; also has audio podcast.

AI with Allie

Newsletter
Turning AI buzzwords into actionable business tactics from a former Amazon AI leader.
Author: Allie K. Miller
Subscription: Free

Techniques and tips for using AI tools.

Artificial Corner

Newsletter
Plain English explanations of AI tools and concepts, with practical how-to guides for using AI and Python, aimed at making AI accessible without a technical background.
Author: Frank Andrade (The PyCoach)
Subscription: Freemium (free weekly; $20/mo or $200/yr for courses, guides, and prompt libraries)

Semi-detailed how-tos — helpful for 0–60 learning.

AI Research & Technical Deep Dives

Interconnects

Newsletter
AI models, training methods, open-source developments, and AI research trajectory.
Author: Nathan Lambert
Subscription: Free

AI researcher’s blog — very technical, but makes sense of broader AI trends.

Featured — Simon Willison’s Newsletter

Newsletter
AI, LLMs, Python, open source, data tools, and practical AI explorations from the co-creator of Django.
Author: Simon Willison
Subscription: Free

Practically technical perspective; tests the technology.

Future of Work & Management

Managing The Future of Work

Podcast
HBS professors talk to leaders grappling with forces reshaping work including AI, automation, and labor markets.
Presenters: Joe Fuller & Bill Kerr

Cold Call

Podcast
Harvard Business School’s legendary case studies distilled into audio with faculty discussing real-world business lessons.
Presenter: Brian Kenny

Product Management & Growth

Lenny’s Newsletter

Newsletter
No-nonsense advice on product management, growth, and career development with an accompanying podcast.
Author: Lenny Rachitsky
Subscription: Freemium (one free post/month; $200/yr for full access)

Lenny’s Podcast

‘Podcast
Interviews with world-class product leaders and growth experts uncovering concrete, actionable advice.
Presenter: Lenny Rachitsky

See also his newsletter.

Source: Ken Calhoon

A pattern I keep encountering in stalled AI projects: the business reports “the AI is outputting wrong results.” Upon closer inspection, the AI is doing exactly what it was asked to do — it was just asked incorrectly. The sequence of steps is wrong, the calculations are fuzzy, the data source is off. The root cause isn’t the model. It’s that nobody documented what “right” looks like before development started.

Teams routinely assume the model will figure out what to do without the hard work of defining success. It won’t. The project pauses, developers sit idle, and the business scrambles to reverse-engineer the very workflow it should have documented upfront. Without a clear process, AI simply scales and automates existing confusion.

Three prerequisites before greenlighting any use case:

1. Documented workflow with clear improvement targets. Map the current process end to end: steps, owners, decision points. Identify specifically where AI adds value. If you can’t write the SOP, the AI can’t execute it.
2. Confirmed and accessible data. Identify the exact systems, fields, and formats required. Verify access and accuracy before engineering begins. Assumed data availability is the fastest path to a stalled project.
3. Golden examples for the engineering team. Provide real, validated outputs that represent “good.” Development compresses dramatically when engineers build against concrete examples rather than iterating through ambiguous feedback loops.

These three items require real investment from the business, not just the technical team. Process definition, data validation, and example curation are business responsibilities.

AI rewards preparation. It punishes ambiguity.

During scoping, the team uncovered critical workflow inefficiencies that had nothing to do with AI. Standard engineering — no neural networks required — fixed the data problem, delivered outsized ROI, and made the downstream human work dramatically easier. What remained was the genuinely hard AI problem: high-nuance exceptions that, at current cost and complexity, didn’t pencil out. Leadership did the math. Most of the gains were already captured. They redirected resources to projects where AI was more feasible and returns were clearer.

The meta-lesson is worth stating directly: sometimes the value of an AI initiative isn’t the model. It’s the rigorous process inspection that building the model requires. If you capture that value before deployment, the win is real; take it.

Too many organizations treat AI initiatives as commitments they can’t walk away from, as if stopping signals weakness or lack of vision. The opposite is true. Dogged pursuit of a marginal use case burns capital and talent that could be deployed where the technology actually fits.

Clear-eyed mid-project assessments separate organizations generating real returns from those accumulating expensive experiments. Sometimes the best AI strategy is a simpler solution.

According to BCG, the top 15% of CEOs are “Trailblazers” transforming AI from a CIO experiment to a business-led operating model. They invest ~1.7% of revenue—allocating 60% specifically to agentic AI—and upskill teams at double the average rate. They advocate simplicity, focus, and pushing through short-term pilot failures. link link

Source: BCG

With Apple officially tapping Google’s Gemini to power Siri, the tech giant is finally optimizing “buy vs. build.” The takeaway? Stop roadmapping the “uninvented.” Instead, lease commodity intelligence and double down on true differentiation—like Apple’s privacy layer—to build the most valuable product. link

The market is pricing this in. SaaS stocks lost ~$300 billion in market cap reflecting investor fears of AI workflow substitution. While the reaction was sharp, Goldman Sachs estimates the Agentic total addressable market will exceed SaaS by 2030. AI threatens to unbundle software, allowing enterprises to customize functionality less expensively. Forbes Goldman Sachs

Source: Goldman Sachs

“Pain is the new moat,” says OpenAI’s Kiriti Badam. Applied anthropologist Mikkel Rasmussen warns that if work feels easy, he gets “nervous.” The real competitive advantage lies in working through the hard-to-solve problems. YouTube Lenny’s Newsletter

Fast for the few

Development is shifting from coding to architecture and review. Anthropic shipped “Claude Cowork” in just 10 days with four engineers using AI to generate 90% of the codebase—then launched it as a desktop tool that gives non-engineers the same power, no terminal required. OpenAI similarly built Sora in 28 days. Product timelines are compressing by orders of magnitude. X Substack Anthropic

Source: Ken Calhoon

“That ship has sailed and good riddance,” said a veteran developer about manual coding in a UC San Diego study. While experts embrace AI agents, the paper reveals they reject “vibe coding,” instead choosing to “control” agents through rigorous planning and supervision to ensure quality. link

Friction for the many

Outside of elite engineering teams, however, general enterprise adoption faces friction. Workday finds AI carries a 37% “rework tax” for correcting low-quality outputs. To reclaim productivity, leaders must guide employees from augmentation to automation—refining prompts, curating reference content, and experimenting with different LLMs to fix instructions rather than results. link

Source: Workday

Microsoft sees this too. According to The Information, the company is reportedly offering large clients tens to hundreds of thousands of dollars to fund employee training and ensure Copilot adoption, concerned that customers aren’t getting enough value from its AI tools. link

Anthropic argues the biggest barriers aren’t model capabilities but data access and context aggregation. Companies with fragmented or siloed data will struggle to unlock sophisticated use cases regardless of technological advancements. link

Security risks compound these adoption challenges. Air Street Press highlights the growing conflict between innovation and safety following OpenClaw’s debut. Cisco deems the tool a “security nightmare,” yet Token Security found 22% of tracked users have installed it on corporate systems. link

The land grab continues

Google says Gmail is “entering the Gemini era,” powered by its Gemini 3 models. New features include an “AI Inbox” to prioritize VIPs and “AI Overviews” that let you ask your inbox questions naturally, turning the email client into a proactive personal assistant. link

OpenAI is expanding on two fronts. It launched ChatGPT Health—a private workspace developed with over 260 physicians that connects with Apple Health and MyFitnessPal—after finding 230 million people already ask its models health questions weekly. And it will test ads in its free and $8/month “Go” tiers to monetize 900M weekly active users, with advertising crucial to fund $1.4 trillion in infrastructure commitments. OpenAI Seeking Alpha Stratechery

The “new front door”

McKinsey warns the “new front door” to the internet is shifting: by 2028, AI could dominate 50% of search activity, influencing $750 billion in commerce. As decision-making moves pre-click, brands relying on traditional SEO risk a 20–50% traffic plunge, necessitating a pivot to AI-native strategies. link

Because AI search is answer-based, brands are scrambling to influence those answers. Ahrefs found the most detailed story wins, even if false. After seeding Reddit and Medium with fake narratives about a made-up brand, they found models like Gemini trusted third-party lies over the official FAQ, making GEO a critical new marketing frontier. link

Source: Ahrefs

On the production side, Colgate-Palmolive worked with Google and BCG to produce video ads 4x to 6x faster using AI-generated imagery. BCG estimates fixing content inefficiencies can save marketers 10% to 30% of their total budget. Google BCG

Source: Menlo Ventures

AI spending is growing fast, but the value equation remains uncertain

Enterprise Gen AI spending hit $37B in 2025, capturing 6% of the SaaS market, according to Menlo Ventures. While $19B went to apps and $18B to infrastructure, architectures remain simple. Prompt engineering and RAG dominate implementation, with only 16% of enterprise deployments qualifying as true autonomous agents. link

While Gen AI spending has grown 3.2x this year, it represents less than 1% of the $4 trillion in 2024 U.S. corporate profits reported by the St. Louis Fed. Even stratospheric returns on AI spend won’t materially impact these aggregate profits for the foreseeable future. link

Microsoft Copilot leverages corporate data in Teams and SharePoint, but users face a steep learning curve, according to The Information. Concerned that customers aren’t seeing enough value to renew subscriptions, Microsoft is now funding employee training, while OpenAI similarly offers resources to bridge the skills gap. link

Companies need “domain owners” who have built a “second muscle” in technology, according to McKinsey. With only 5% of senior leaders having held technical roles, these executives are vital for aligning data, platforms, and engineering talent to deliver successful AI transformations. link

Source: Anthropic

Who’s capturing that spend? Anthropic and open-weight models surge ahead

Menlo Ventures also reports Anthropic has unseated OpenAI as the top enterprise model provider, capturing 40% of spend versus OpenAI’s 27%. Google also grew to 21%. Anthropic’s surge is powered by a dominant 54% share of the coding market and the success of Claude Code. link

Anthropic also reports its engineers primarily use Claude for debugging and code understanding. Staff use the tool in 60% of their work, estimating a 50% productivity boost. Additionally, 27% of Claude-assisted work involves “nice-to-have” tasks that otherwise wouldn’t be completed. link

A new analysis of the Hugging Face ecosystem details a “fundamental rebalancing” of AI power. As US corporate dominance wanes, Chinese developers have surged ahead, with DeepSeek and Qwen driving China’s download share to 17.1%, surpassing the collective output of American model builders for the first time. link

Exemplifying the above trend, Pinterest and Airbnb have fine-tuned open-weight models from Qwen and DeepSeek to match frontier performance. According to Applied AI, these models handle multi-modal jobs and customer support queries at less than 10% of the cost. link link

Despite the hype, most implementations remain focused on the basics

OpenAI reports that technology, healthcare, and manufacturing are the fastest-growing AI adopters, with the tech sector surging 11x year-over-year. While professional services and finance continue to lead in absolute scale, a widening divide shows “frontier” firms now send twice as many messages per seat as the median enterprise. link

Source: Anthropic

While most usage remains in “predictable areas,” Anthropic highlights N26 and Thomson Reuters leveraging Amazon Bedrock for compliance and security. Anthropic calculated a 1% increase in input context length correlates with a 0.38% increase in output quality, demonstrating that robust, selective data aggregation, not just model power, is critical for success. link

Come together, right now

AWS and Google announced a jointly developed service providing high-speed networking links between their competing clouds. This allows enterprises to host data in their preferred cloud while utilizing models from other providers, reducing connection setup times from weeks to minutes. link

The Linux Foundation is organizing the Agentic Artificial Intelligence Foundation to develop shared open-source standards, according to The Information. Initial technical contributions include OpenAI’s Agents.md, Block’s Goose, and Anthropic’s MCP. OpenAI, Google, Microsoft, and Anthropic have joined the group to ensure agent ecosystems can effectively connect. link