Clouds and AI labs are signing massive, long-term deals to lock up compute and revenue just as open-source models get good and cheap enough to challenge their pricing power.
At the same time, a string of ugly breaches and governance workarounds is making it much less obvious which AI platforms can actually be trusted with serious workloads.
Key Events
/Amazon deepened its partnership with Anthropic, giving the lab access to up to 5GW of AWS compute capacity for AI training.
/Tim Cook is stepping down as Apple CEO, with hardware chief John Ternus taking over while Cook becomes Executive Chairman.
/Google Cloud Q4 revenue reached $17.7B on 48% year-on-year growth, with its backlog projected to hit $240B by 2025.
/AI app builder Lovable disclosed a mass data breach exposing all projects created before November 2025 due to an unresolved bug.
/Vercel reported a security breach after an employee granted an AI tool unrestricted access to Google Workspace, with attackers seeking $2M for the stolen data.
Report
The real game this month is who owns the AI pipes: hyperscalers, labs, or open models. Everything else — Apple’s regime change, the breaches, even the NSA drama — are just consequences of that fight over compute, data, and trust.
the cloud–AI land grab
Amazon is tying itself tightly to Anthropic, giving the lab up to 5GW of AWS compute capacity for model training, with nearly 1GW expected online by 2026.
Anthropic has committed roughly $100B of future spend on AWS under the same arrangement.Google Cloud is riding the same wave, with Q4 revenue hitting $17.7B on 48% year-on-year growth from AI-heavy workloads.
Its contracted backlog is projected to reach $240B by 2025, effectively pre-selling a huge chunk of future compute and storage. In the US, only about 5GW of the roughly 140 planned datacenter capacity projects are actually under construction, so hyperscalers are pre-allocating a scarce resource rather than a commodity.
open models are eating proprietary margins
Open-source model Kimi K2.6 scored 58.6 on the SWE-Bench Pro coding benchmark, surpassing flagship closed models from Anthropic and OpenAI in published comparisons.
Developers also report Kimi sustaining more than 4,000 tool calls over 12 hours on long-horizon coding tasks, which is the kind of behavior proprietary agents usually sell as premium.
On price, Anthropic users say Kimi is roughly 76% cheaper than Claude Opus 4.7 on a per-token basis. Separate analyses find that open-source models can be up to 65% cheaper than proprietary ones for many workloads, narrowing the room for closed vendors to charge a performance premium.
At the edge, local LLMs running on Macs with 32–64GB of RAM already deliver performance comparable to Claude Sonnet-class models for many tasks, even as community voices argue that today’s benchmarks don’t capture long-run reliability.
AI adoption without productivity
A Gallup survey reports that about half of employed Americans now use AI at work in some form. Among low-income Americans, 32% say they use AI as a substitute for doctor visits they can’t afford, despite broad distrust of AI in healthcare.
In a large sample of AI-generated software projects, only 1% passed production-readiness checks when reviewers looked at operational hygiene. Thousands of CEOs still tell surveyors that AI has had no material impact on either employment levels or measured productivity inside their firms.
Engineers echo the pattern: AI coding tools on platforms like GitHub are widely adopted, but many teams report little improvement in real project outcomes and note that LLMs shine only in human-in-the-loop workflows, not deterministic tasks.
platform and silicon reset at apple
Apple is handing the CEO job from Tim Cook to longtime hardware chief John Ternus just as it hits its 50th anniversary, with Cook moving to Executive Chairman.
Ternus built his career in hardware engineering, including early VR headsets, and investors have nudged the stock up on expectations that he will prioritize device innovation over incremental services tweaks.
In parallel, Apple has confirmed that macOS 27 will drop support for Intel processors entirely, accelerating the shift of its ecosystem onto Apple Silicon.
From 2027, EU rules will require all phones sold in the bloc to have replaceable batteries, directly challenging Apple’s sealed-device design and especially its high-margin low-end hardware.
This is also the company that forced eSIM adoption despite initial carrier resistance, so a hardware-focused CEO under growing regulatory pressure points to more aggressive moves in radios, form factors, and platform control.
AI security and trust as a breaking point
AI startup Lovable disclosed that a long-standing bug had let anyone with a free account access source code and customer data for all projects created before November 2025.
AI cloud company Vercel reported that an employee granted an AI tool unrestricted access to Google Workspace, enabling attackers to breach internal systems and try to sell stolen data for $2M on underground forums.
On the open-source side, researchers flagged an uncensored cybersecurity model on Hugging Face that ships potential exploit code and has quietly logged hundreds of downloads despite almost no visible review.
Inside government, the NSA is reportedly using Anthropic’s Mythos model even though the Pentagon has blacklisted the vendor as a supply-chain risk, revealing a willingness to bend rules when the models are operationally useful.
Meanwhile in consumer social, French prosecutors are investigating Elon Musk’s X for possible distribution of child abuse imagery and deepfakes, showing how AI-boosted content can turn directly into criminal liability for platforms.
What This Means
AI is concentrating power in a handful of clouds and labs just as open models compress pricing and security failures expose how leaky the stack really is. The real decision is where you deliberately trade away control—for compute, models, or platforms—and where you insist on keeping leverage even if it slows you down.
On Watch
/Rumors that Anthropic may be eyeing an Atlassian acquisition, alongside Atlassian’s shift to default data collection for AI training, signal that enterprise workflow data is becoming a first-class acquisition target.
/China’s domestic accelerators and supply chain are ramping fast, with local chips projected to capture 41% of the AI server market by 2025 and claims that DeepSeek V4 can deliver 35x faster inference without NVIDIA GPUs.
/New EU and US pushes for age‑verification at the OS and app level, plus World ID’s expansion and Zoom’s human-ID partnership, point toward an emerging regulated digital identity layer that could reshape onboarding and payments.
Interesting
/AI inference costs are nearing 10% of engineering headcount costs, highlighting uneven adoption despite enthusiasm.
/The physical infrastructure for U.S. datacenter projects is significantly sourced from China, highlighting interdependence.
/The HiFloat4 training format developed by Huawei outperformed the Western-developed MXFP4 format in a recent chip bakeoff, indicating advancements in training efficiency.
/Maryland's law against 'surveillance pricing' targets companies like Amazon for price discrimination based on user data.
/Concerns are emerging about NVIDIA's market position as rivals develop custom silicon solutions that could reduce costs and challenge its dominance.
We processed 10,000+ comments and posts to generate this report.
AI-generated content. Verify critical information independently.
/Amazon deepened its partnership with Anthropic, giving the lab access to up to 5GW of AWS compute capacity for AI training.
/Tim Cook is stepping down as Apple CEO, with hardware chief John Ternus taking over while Cook becomes Executive Chairman.
/Google Cloud Q4 revenue reached $17.7B on 48% year-on-year growth, with its backlog projected to hit $240B by 2025.
/AI app builder Lovable disclosed a mass data breach exposing all projects created before November 2025 due to an unresolved bug.
/Vercel reported a security breach after an employee granted an AI tool unrestricted access to Google Workspace, with attackers seeking $2M for the stolen data.
On Watch
/Rumors that Anthropic may be eyeing an Atlassian acquisition, alongside Atlassian’s shift to default data collection for AI training, signal that enterprise workflow data is becoming a first-class acquisition target.
/China’s domestic accelerators and supply chain are ramping fast, with local chips projected to capture 41% of the AI server market by 2025 and claims that DeepSeek V4 can deliver 35x faster inference without NVIDIA GPUs.
/New EU and US pushes for age‑verification at the OS and app level, plus World ID’s expansion and Zoom’s human-ID partnership, point toward an emerging regulated digital identity layer that could reshape onboarding and payments.
Interesting
/AI inference costs are nearing 10% of engineering headcount costs, highlighting uneven adoption despite enthusiasm.
/The physical infrastructure for U.S. datacenter projects is significantly sourced from China, highlighting interdependence.
/The HiFloat4 training format developed by Huawei outperformed the Western-developed MXFP4 format in a recent chip bakeoff, indicating advancements in training efficiency.
/Maryland's law against 'surveillance pricing' targets companies like Amazon for price discrimination based on user data.
/Concerns are emerging about NVIDIA's market position as rivals develop custom silicon solutions that could reduce costs and challenge its dominance.