DistillBy MirusRequest early access
Research, distilled

Not a chatbot answer. A finished artifact.

Distill produces deep, citation-verified, rubric-driven research documents for executives who can’t afford to be wrong.

Inline citations
Every claim traceable to its source
Rubric-driven
Structured to a defined standard of rigor
Source quality
Primary, peer-reviewed, or modeled — labeled
Gap analysis
What we couldn't confirm, and why
INF-2026-0512RUBRIC: TECH LANDSCAPE · 14 PP
Executive summary

The inference bottleneck: where capacity meets cost.

Compute supply for foundation-model inference is repricing along three vectors at once — power, memory bandwidth, and packaging. Each constraint resolves on a different timeline, and the order they unlock will determine which providers absorb the marginal workload through 2027.

3.4×
YoY inference demand growth, blended
22%
Price gap, leading vs prior gen
120d
Median packaging-capacity lead time
Structured narrative

Inference demand is projected to grow [1] at roughly 3.4× year-over-year through 2027, while accelerator production remains throttled by HBM yield and CoWoS packaging capacity. The result is a sustained gap between booked and deliverable compute that flows through to enterprise pricing.

Demand is currently constrained by packaging, not silicon. The constraint that resolves first will decide which vendor lands the marginal workload.
IEA — World Energy Outlook 2025PrimaryTSMC FY25 Q4 capacity disclosurePrimarySemiAnalysis HBM model, Mar 2026ModeledBain — Generative AI in the Enterprise 2026Peer-reviewed
Gap analysisProvider-specific allocation share past Q4 2026 could not be confirmed — disclosure is inconsistent across vendors and counter-party NDAs limit cross-checking.
Follow-on questions
  • →Which constraint will resolve first — power, memory, or packaging?
  • →How does HBM3e ramp pacing change the 2027 forecast?
  • →What does a 6-month packaging delay do to enterprise pricing?
Sources: IEA · TSMC · SemiAnalysis · Bain · 41 citations audited
The problem

Research tools give you information. Distill gives you clarity.

What you have now
What it costs you
What Distill does instead
ChatGPT / Perplexity

Confident answers with no audit trail. You can’t tell what’s sourced.

Every claim labeled with its source and tier — primary, peer-reviewed, or modeled.

A junior analyst

Slow, expensive, and inconsistent. Quality depends on who you got.

A defined rubric and a citation audit — same standard, every time.

Nothing

You decide on instinct or rumor. The bad outcomes are quiet but compound.

A finished research artifact, in 24 hours or less, written for executives.

What is a Distillation

Not a chatbot response. A finished artifact.

Six fixed parts. Same shape every time. The reader knows where to look.

  1. 01

    Executive summary

    The decision-ready version, set up so a busy reader can act on the first page.

  2. 02

    Structured narrative

    A defined arc through the question — claims, evidence, and trade-offs in order.

  3. 03

    Inline citations

    Every load-bearing claim carries a numbered citation traceable to its source.

  4. 04

    Source quality indicators

    Primary, peer-reviewed, or modeled — labeled on the chip itself.

  5. 05

    Gap analysis

    What we couldn’t confirm, and why — never glossed over.

  6. 06

    Follow-on questions

    The next questions worth asking, sequenced by what we learned.

INF-2026-0512RUBRIC: TECH LANDSCAPE · 14 PP
Executive summary

The inference bottleneck: where capacity meets cost.

Compute supply for foundation-model inference is repricing along three vectors at once — power, memory bandwidth, and packaging. Each constraint resolves on a different timeline, and the order they unlock will determine which providers absorb the marginal workload through 2027.

3.4×
YoY inference demand growth, blended
22%
Price gap, leading vs prior gen
120d
Median packaging-capacity lead time
Structured narrative

Inference demand is projected to grow [1] at roughly 3.4× year-over-year through 2027, while accelerator production remains throttled by HBM yield and CoWoS packaging capacity. The result is a sustained gap between booked and deliverable compute that flows through to enterprise pricing.

Hyperscalers have responded by [2] shifting to multi-vendor inference paths and pre-buying capacity from second-tier silicon providers. The shift narrows the price gap between bleeding-edge and one-generation-back hardware to less than 22%.

The constraint that resolves first — power, memory, or packaging — will decide which vendor lands the marginal hyperscaler workload.
Inline citations
IEA — World Energy Outlook 2025PrimaryTSMC FY25 Q4 capacity disclosurePrimarySemiAnalysis HBM model, Mar 2026ModeledBain — Generative AI in the Enterprise 2026Peer-reviewed
Source quality indicators

Each citation carries its tier — primary, peer-reviewed, or modeled — visible on the chip itself. Modeled claims aren’t suppressed; they’re labeled.

Gap analysisProvider-specific allocation share past Q4 2026 could not be confirmed — disclosure is inconsistent across vendors and counter-party NDAs limit cross-checking.
Follow-on questions
  • →Which constraint will resolve first — power, memory, or packaging?
  • →How does HBM3e ramp pacing change the 2027 forecast?
  • →What does a 6-month packaging delay do to enterprise pricing?
Sources: IEA · TSMC · SemiAnalysis · Bain · 41 citations audited
How it works

Layered research. Transparent process. Verified output.

1

You ask

A clear question, a decision context, and your standard of rigor.

2

Rubric selection

A rubric chooses the shape of the answer before research begins.

3

Research pipeline

Layered retrieval and synthesis run against verified sources.

4

Intermediate artifacts

Drafts, evidence packs, and counter-arguments along the way.

5

Citation audit

Every claim is matched back to its source — or removed.

6

Your Distillation

A finished, citation-verified document, ready to act on.

The rubric advantage

Methodology shouldn’t be left to chance.

Every Distillation is written against a rubric — a defined standard for what counts as evidence, how it’s weighed, and what the finished document must contain. Choose from the system rubrics, or define your own.

Or define your own — for your industry, your decision type, your standard of rigor.

  • RUB-01Company InvestigationSystem
  • RUB-02Investment Due DiligenceSystem
  • RUB-03Competitive AnalysisSystem
  • RUB-04Technology LandscapeSystem
  • RUB-05Market DynamicsSystem
  • RUB-06Your own rubricCustom
Privacy

Your research stays yours.

01

We do not train on your work. Ever.

Your prompts, your sources, and your Distillations are excluded from any model training. Contractually.

02

Distillations are private by default.

Only you can see them until you choose to share. No public index, no observability into your decisions.

03

Sharing is explicit, and controlled by you.

Share to named recipients on revocable links — every access logged, every share auditable.

Early access

Built for executives who can’t afford to be wrong.

The research you need. The rigor you’d demand from your best analyst. The privacy you require for your most sensitive decisions.

Early access · Invitation reviewed within 48 hours