7 min read · 11 sources

Google spent the week turning its data cloud into a federation layer rather than a destination. The centrepiece is a “borderless Lakehouse” on Apache Iceberg, cross-cloud interconnect and caching for AWS and Azure now generally available with zero egress fees, and catalog federation into Databricks Unity, Snowflake Horizon and AWS Glue.

That is a direct shot at the copy-everything-into-one-warehouse model that has defined the last decade of analytics. Google is saying the data can stay where it is, including in its competitors’ catalogs, and you query it through Gemini-assisted tooling instead.

The rest of the day is a study in what AI actually costs. Microsoft is capping Copilot Chat without telling anyone the numbers, Apple is capping token budgets with numbers it also hasn’t confirmed, and Binance lost $6.3 billion to an attack that started with a developer installing the wrong Homebrew formula.

Binance lost $6.3 billion after attackers got in through a tampered Homebrew install of an AI coding tool and a fake interview test.

Google Wants the Lakehouse to Be Borderless

Iceberg is the open table format that lets different engines read the same files on object storage, and it has become the neutral ground in the warehouse wars. Google’s pitch is that you no longer copy those tables into BigQuery to query them. Catalog federation means BigQuery can read metadata from Databricks Unity, Snowflake Horizon and AWS Glue without a migration project, and cross-cloud interconnect plus local caching means the bytes come over a managed path instead of a public-internet hop.

The caching layer is the part operators should care about. Zero egress fees are a pricing decision; local caching is a latency decision, and it changes the shape of a query plan when half your joins read from S3 in another region. Google also added Gemini natural-language querying and Knowledge Catalog context curation, aimed at agents that need to find the right table without a human writing the join.

What it replaces is the pipeline sprawl: CDC jobs, nightly extracts, and a serving copy per cloud. What it adds is a dependency on catalog APIs you don’t control, and a governance problem that now spans systems owned by three other vendors. Federated access is not the same as federated permissions, and that gap is where the incidents will come from.

Microsoft's Copilot Chat Caps Are Secret, and That's the Point

Microsoft is expanding usage limits for commercial Copilot Chat users who don’t hold a paid Microsoft 365 Copilot licence. It won’t disclose what the limits are, and it won’t let administrators adjust them. The rollout started November 2 for smaller organisations and reaches tenants with more than 2,000 seats on November 9.

For anyone running a helpdesk, this is the worst possible shape for a quota. Users hit an invisible wall, admins can’t see it coming, and there’s no throttle to tune. Capacity planning becomes a matter of waiting for the tickets. Compare that with Apple, which at least leaked a number.

Apple Cut the Daily AI Token Allowance From $300 to $200

Apple is reportedly tightening employee AI budgets, dropping daily allowances from $300 to $200 and planning caps of $100 for engineers and $25 for other staff. The limits apply to third-party models including Anthropic’s Claude and OpenAI’s ChatGPT, and some teams are being pushed toward cheaper or locally hosted models. Apple hasn’t confirmed any of it.

If the numbers hold, the interesting part isn’t the cut - it’s the tiering. A $25/day ceiling for non-engineers is enough for chat and drafting; a $100 ceiling for engineers is not enough for agentic coding loops that burn tokens on test runs. That pushes exactly the workloads that benefit most from big models back toward smaller local ones, which is a capability decision disguised as a finance one.

AI Made Bug-Finding Cheap. Patching Is Still Two Months and Four Engineers.

Source: hitesh.in ↗

A PHP vulnerability called PHP2Shell sat in the backlog as a low-priority out-of-memory bug. AI-assisted researchers turned it into remote code execution. The fix took two months and four engineers, according to a write-up on the asymmetry.

The argument is straightforward: AI collapses the cost of discovery, and remediation still moves at human speed. Every “won’t fix” in the tracker was priced against the old cost of finding it. The suggested response is unglamorous - audit deferred security tickets, sunset internet-facing systems nobody uses, and actually review the accepted-risk register on a schedule instead of at audit time.

Binance's $6.3 Billion Lesson: the Boring Layer Is the Attack Surface

Source: techtrenches.dev ↗

Binance lost $6.3 billion after attackers compromised developer laptops with a tampered Homebrew install of an AI coding tool and a fake interview test, then abused an internal AI tool to obtain a container-deployment token, per Tech Trenches.

Nothing in that chain is exotic. It’s a supply-chain compromise, a social-engineering lure, and an internal service with more privilege than it needed. The AI tooling is what made the internal service interesting to attackers in the first place. If your coding assistants and internal agents can reach a deployment token, they are production infrastructure and should be reviewed like it.

MariaDB 2027 Puts Agent Memory Inside the Transactional Database

Source: mariadb.com ↗

MariaDB shipped Enterprise Platform 2027, a single platform combining transactions, real-time analytics and in-memory speed with agent-oriented features: Agent Memory, Semantic Cache, AI Retrieval and AI Plugins. The headline benchmarks are 56% higher peak throughput than MariaDB 11.8 and 25% faster than MySQL 9.7.2, plus BYOC deployment, FIPS 140-3 compliance and active-active multi-region clustering on MariaDB Raft Cluster.

Semantic Cache is the one to watch on the bill. Caching LLM responses by meaning rather than exact prompt cuts both latency and token spend, and doing it next to the transactional data avoids a second consistency problem. The trade-off is the usual one: you now have a cache whose invalidation semantics you have to reason about.

Confluent's Lightning Tables Put SQL Directly on Kafka

Source: confluent.io ↗

Confluent made Lightning Tables generally available on AWS, letting applications run live SQL and API queries over continuously updated Kafka data. It’s the anchor of the company’s “Streamhouse” architecture, which bundles streaming, processing, governance and querying into one stack.

The pattern it removes is familiar: stream to Kafka, CDC into a serving store, then query the copy and hope the lag is small enough. Querying the stream directly kills the second write and the staleness window. For agents that need current business state, that’s the difference between an answer and a plausible answer.

EXACT: TDD for Agents Without the Vibe-Coded Rot

Source: coding-is-like-cooking.info ↗

A write-up describes EXACT, a workflow that bolts test-driven development onto AI coding agents. It leans on example mapping to build a shared human model, converts that into natural-language test lists, and uses predictive TDD to drive the agent. The reported experiments show better correctness and code quality than unguided vibe coding or plain TDD prompts.

The mechanism is unremarkable and that’s why it works: concrete examples beat abstract specs, and an agent that must satisfy a failing test has less room to invent an API.

Google's London Generators Failed on a Voltage Sensor

Source: assets.theregister.com ↗

Google published a postmortem for its October London outage, blaming a voltage-sensing flaw and a controller misconfiguration that stopped the backup generators from starting. One power feed was already under repair when the second failed, and the battery plant ran out before power came back.

The failure mode is the one every DR plan should be tested against: redundancy that shares a sensing or control path isn’t redundancy. Two feeds and a generator farm are one system if a single misconfigured controller decides whether the generators start.

AlloyDB Express Spins Up Postgres in Under Two Seconds

Source: cloud.google.com ↗

Google previewed AlloyDB Express, a Postgres service that provisions in under two seconds, auto-pauses after 15 minutes of idle time, and bundles vector, text and hybrid search plus access to Gemini models. There’s a free tier, pay-as-you-go pricing, and a one-click upgrade to full AlloyDB with HA and DR.

The auto-pause is the cost lever. Prototype databases sit idle most of their lives, and a 15-minute scale-to-zero turns a fixed monthly line item into noise.

The Briefs: PyTorch 2.15, Meta's Conductor, and a £20B Biomass Datacenter

Source: diginomica.com ↗

PyTorch 2.15 is aimed squarely at the “one bad GPU” problem, making large-scale training more resilient when a single accelerator in a big job misbehaves - the kind of failure that otherwise costs you a restart and a day. Meta bought Conductor, an agent orchestration startup, giving it an exit. On-prem server sales for HPC and AI are up 20% this year, and the reporting is blunt that the revenue spike is as much about rising costs as about more units. The Windows AI PC push has run into RAM prices, leaving customers sweating assets on tight budgets. Drax wants to build a £20 billion datacenter at its biomass power station to feed Britain’s AI habit, but needs planning permission, government approval and, rather importantly, a customer. On the infosec side, attackers are impersonating Signal support to run phishing campaigns, and the US took down a set of Iranian propaganda sites. Oracle’s Hari Sankar, meanwhile, argues the next ERP generation is about agents executing outcomes like automated account reconciliation rather than surfacing insights - which means the integration surface you expose to those agents is now the product.