Amazon's $1B charm offensive meets the water question it calls a myth
Amazon said on 2 October it will spend more than $1 billion over five years on community programs around its datacenters, a push that includes arguing local water fears are overblown.
The announcement came in a company blog post. The Register's Tobias Mann read it as a "whataboutism-packed" attempt to win over residents who do not want the bit barns in their backyards. Amazon framed the spend as education, job training, local infrastructure and sustainability work in datacenter host communities.
The water argument is the part worth pulling out. Amazon says its facilities use about 13,000 gallons a day, the equivalent of 42 US households, against a claimed industry figure of five million gallons a day for a typical datacenter. It also argues citizens should worry more about golf courses, which it says use 200 times more water than all of Amazon's datacenters.
The Register's rebuttal is that the comparison skips the power side. Datacenters can be built to use little water, but those designs often burn more electricity, and most gas, nuclear and coal plants use steam to turn a turbine, which means water. The blog post also warns that more than 100 datacenter moratoriums are under consideration across the country and says foreign actors are seeding misinformation.
Anthropic's bug hunter, and the water it never mentions
The compute behind these arguments is the same compute now being pointed at security work. On Wednesday, Horizon3 researcher Zach Hanley said he used Anthropic's unreleased Mythos model to find CVE-2026-61500, a critical authentication bypass in Rejetto HTTP File Server that can lead to full admin access and remote code execution, The Register reported. Users should move to v3.2.1 or later.
Exploitation followed fast. VulnCheck's Patrick Garrity said on Thursday that his canaries caught an actor in China hitting real vulnerable hosts in the US and Japan, and by Friday he had seen four more hits from two US IP addresses in the same subnet that "appear to be coming from a proxy." Garrity's tracker counts 286 CVEs from Mythos and Project Glasswing, Anthropic's partner program, with only two exploited in the wild so far.
"Today we have seen four hits," Garrity told The Register on Friday.
Anthropic is also arguing that open-weight rivals are catching up. A new company report says Zhipu AI's GLM-5.3 can generate exploits at close to Mythos levels, per Tom's Hardware: 50 end-to-end Chrome exploits in 410 runs against Mythos's 56, and a 4% success rate on a control-flow hijack benchmark against 6%. Kimi K3 and DeepSeek V4.1 Flash scored 0%. Anthropic says guardrails fell to a prefilled-thinking trick 92% of the time.
What the smaller releases tell us
Not every model launch is a frontier event. Aleph Alpha published Kolibri on 3 October, a 78B-parameter mixture-of-experts model with 3B active parameters, a 1M-token context and Apache 2.0 weights, pitched at sovereign work in public administration, industrials and aerospace. The same day, llama.cpp added a /v1/systemone endpoint for decision models that score options instead of writing text, with the smallest, Julia-1, answering in 3 ms on an RTX PRO 6000 according to the Hugging Face write-up.
Elsewhere the week produced a 118M-parameter prompt-injection classifier called Shield Small, released 2 October under int8 ONNX at 118 MB and runnable on a CPU, and a 10-player CS:GO dataset from Vortech with 3,958 hours of 5v5 matches and 195M state-to-action steps, licensed CC BY-NC 4.0.
Google, meanwhile, is trimming what free users get. Its help page says that from 9 October, accounts without an AI subscription default to automatic model selection, with most prompts routed to Flash-Lite, and usage limits move to a compute-based system that refreshes every five hours.
None of these posts quantify water. That is the gap Amazon's critics keep pointing at, and a $1 billion community fund does not fill it.
Sources
10- 01Shut up and take our money: Amazon plows $1B+ into quashing datacenter dissentEN
- 02Anthropic's super bug-hunting model Mythos is hardcore good at math, as latest vuln under attack showsEN
- 03Anthropic claims popular Chinese AI model has Mythos-class hacking abilitiesEN
- 04Kolibri Has Landed: A Sovereign Open-Weight ModelEN
- 05New in Llama.cpp: Decision ModelsEN
- 06Shield: A 118M model for detecting prompt injections and jailbreaksEN
- 07Open 10-player CS:GO dataset for multiplayer world modelsEN
- 08Changes to Gemini model access and limitsEN
- 09OpenAI alerts 100+ orgs that its 'misaligned models' attempted to break in or worseEN
- 10Ideogram 4.5: The most precise edit modelEN
All figures and quotations in this text come from the sources listed below.
Content prepared by the editorial team with AI assistance.
Comments
0- No comments yet — be the first.