Science & Analytics

The Scientist

The Signal

Hyperscalers hold advance deposits on nearly all of 2027's global DRAM output.

The 500% price jump over twelve months is the smaller number. The thing it doesn't tell you is that prepaid capacity turns a cost line into a scheduling constraint: memory-heavy instances and on-prem upgrades may simply be unbookable at renewal, whatever the budget says. The ANN indexes, hot feature caches, and wide Ray executors you sized when bytes were cheap are where that lands, and Nvidia is still planning for plentiful chips.

In Play

  1. Reported Effects Without a Variance Model

    AI News relayed a finding that floating-point accumulation order and sharding layout produce pretraining run-to-run variance nearly as large as seed and data order. Pivot 5 separately decomposed Anthropic's 14-of-15 protein result into 354 binders from 1,320 designs, a 26.8% per-design rate. Both point the same way: a reported effect with no model of the system's own variance is not an experiment. The numerics claim arrives as a relayed summary without ablation tables, so test it before you cite it.

  2. Bytes Inflate While Accelerators Are Promised Cheap

    Memory prices are up 500% in 12 months, with per-unit RAM back at 2007 levels and hyperscalers reportedly holding advance deposits on almost all 2027 global DRAM capacity, per AI News. Every RAM-bound line item you own was sized when bytes were cheap: in-memory ANN indexes, feature-store hot caches, wide Spark and Ray executors. Bloomberg separately reports Nvidia building demand for an era 'when chips will be plentiful' — so accelerator hours may deflate while the bytes around them inflate.

  3. Agent Cost Lives Inside the Orchestration Graph

    Gartner projects inference cost per agent workflow rising more than fivefold through 2028 even as per-token prices fall, per Computerworld. Daily Dose of Data Science reports an open harness matching a managed one on DevRev's Enterprise-Bench at roughly a third of the tokens and 40% fewer model calls. Your spend is set by trajectory design, not by a vendor price list. Both figures are single-run and vendor-produced, so the reproduction is the actual work.

  4. Exploited Flaws in the Systems That Hold Your Lineage

    CISA added four actively exploited flaws to its KEV catalog, including Microsoft SharePoint and VMware vCenter, per The Hacker News. CSO reporting describes an unauthenticated GitLab flaw permitting repository modification and deletion with no user interaction. Those are the systems holding your RAG corpus, your cluster hypervisor and your DVC pointers and training DAGs. A rewritten commit still hash-matches its lineage, so your reproducibility claim inherits your source control's integrity.

  5. Multi-Agent Contention Is Unmeasured

    A study instrumenting 1,902 multi-agent coding runs as temporal networks found direct messaging grows near-quadratically with team size, and that replacing repeated one-to-one messages with shared files cut output tokens about 42% at eight agents, per AI News. Naming a coordinator did not reliably improve outcomes. Risky Business separately reports Anthropic observing agents on identical tasks interfering with each other, but supplies no sample size or task specification, so treat that as a reproduction target.

Deep Dives

  1. Your Noise Floor Lives in the Sharding Config

    Three headline results were reported with no model of the variance in the system that produced them, and each one has a cheap test that settles the question.

    Why a sharding change is a treatment, not a config Floating-point addition is not associative . The order in which partial sums get reduced changes the result in the low bits. Move a run from 8-way to 16-way tensor parallelism…

    3 action items

  2. Bytes Got Expensive While Everyone Watched FLOPs

    Every in-memory index and cache tier you own was sized under an assumption that just inverted, and the compression trade you skipped is now the highest-return work on the board.

    The capacity fact matters more than the price fact The 500% is not the planning input. The planning input is the report that hyperscalers hold advance deposits on nearly all of 2027's global DRAM production , per AI News. That…

    3 action items

  3. Cost Per Resolved Task Is the Column Your Harness Doesn't Have

    Agent spend grows superlinearly inside your own orchestration graph, which makes the largest controllable variable in the stack invisible to any scorecard reporting pass rate alone.

    The mechanism, with arithmetic Daily Dose of Data Science works the example that makes the growth rate visible. An agent queries a CRM at step four and gets back 400 rows . Those rows sit in conversation history. By step…

    3 action items

  4. Your Reproducibility Guarantee Inherits Your Source Control's Integrity

    A rewritten commit and a stolen browser cookie both defeat provenance controls without triggering a single failing test anywhere in your training pipeline.

    The failure mode is lineage poisoning, not code loss An attacker who can modify a repository with no credentials and no user interaction has better options than deletion: rewriting a feature transformation, repointing DVC or LFS at a poisoned dataset,…

    3 action items

The edition continues

Take the signal into the room.

Sign up or log in to read all 4 deep dives in full, plus the final take.

Read the full edition

Continue with LinkedIn