Independent work sample by Chris McCool · not an official Etched site · not affiliated with or endorsed by Etched
Etched GTM Strategy
01 · Etched GTM Strategy · market-development perspective

From purpose-built silicon to scalable inference economics.

A commercial AI infrastructure opportunity across neoclouds, model developers, financial services, enterprise platforms, and potential government partners.

01

Workload advantage

Purpose-built inference systems could compete where supported workloads are high-volume and latency-sensitive; prove the advantage on the buyer's models.

02

Customer economics

The buyer cares about cost per output token, tokens per MW, and capacity delivered — measured on their models.

03

Route to market

Direct lighthouse accounts, neocloud capacity partners, and integrator-led regulated channels.

02 · Why inference, why now

Training develops the model. Inference runs it when someone uses it.

Training and retraining are development workloads. Inference is the recurring, usage-driven compute of serving answers; demand and total cost depend on adoption, model choice and efficiency.
Industry context

Token demand

Each prompt and response consumes compute. As AI is embedded in products, output-token volume becomes an operating line item.

Industry context

Reasoning models

Models that reason before answering can generate additional tokens per task, raising serving compute; the mix of training and inference cost remains workload-dependent.

Industry context

Agentic workloads

Agents call models repeatedly in loops — planning, tool use, verification — multiplying tokens per business outcome.

Key uncertainty: the price-per-token paradox

Falling per-token prices may expand total inference spend (more use cases become viable) — or may compress it if efficiency gains outpace demand. This site makes no market-size forecast; the GTM case must be proven account by account.

03 · What Etched actually sells

Chips, systems, software — delivered as racks.

Current company materials describe frontier inference clusters using Low Voltage Inference (LVI), Cluster Scale Memory (CSM) and an HBM/SRAM hybrid design — not simply the transformer-only Sohu concept reported in 2024. Etched's August release claims support for massive MoE models and non-transformer designs; this is not independent validation or a claim of universal compatibility. Verify the supported envelope at Etched's June 30, 2026 architecture announcement and etched.com; no benchmark numbers are asserted here.

Silicon

Company-described LVI compute and an HBM/SRAM memory design; exact model compatibility must be checked.

Systems

Frontier inference racks/clusters; the first rack delivery is confirmed by the Jane Street announcement.

Software

Serving stack and kernels; company materials describe fleet software and kernel agents as active work.

  1. Jun 30 2026 Sourced

    Etched describes frontier inference clusters using Low Voltage Inference and Cluster Scale Memory, and says production is underway to fulfill over $1B in customer contracts. Company claim; terms and recognized revenue are not disclosed here. Source ↗

  2. By Aug 18 2026 Sourced

    Etched confirms its first rack shipment to Jane Street. This dates the public confirmation, not the exact shipment day or a scaled rollout. Source ↗

  3. Aug 18 2026 Sourced

    Etched announces $700M raise led by Jane Street; Jane Street quoted as having its own rack running in its data center. Source ↗

  4. Aug 18 2026 Sourced

    Press release states >$1B in customer contracts across AI companies and clouds (company claim, terms not public). Source ↗

  5. Next Hypothesis

    Scaled multi-customer production deployment — not yet independently evidenced here.

04 · Target account map

Where purpose-built inference could land first.

Unqualified prospect hypotheses — NOT Etched customers. Company names illustrate segment archetypes only; no relationship, interest or evaluation is implied.
SourcedPublicly identified customer / reference: Jane Street

Named by Etched as its first customer, with a rack running in Jane Street's data center. Etched announcement ↗

Neocloud / GPU hosting

Hypothesis

CoreWeave, Crusoe, Lambda

Hardware buyer vs API consumer
Potential direct hardware buyers or capacity lessees; establish whether each owns its fleet or rents capacity before qualifying.
Top inference workloads
Hosted open-weight LLM serving, reserved inference capacity, MoE model hosting.
GTM route
Direct strategic sale; co-marketing of differentiated inference SKUs.
Qualification questions
What share of fleet revenue is inference? Which models do tenants run? Power headroom per site?
Proof of ROI needed
Revenue per MW and per rack versus GPU baseline at matched latency SLOs.
Risks
Deep NVIDIA relationships and financing tied to GPU collateral.

Frontier & model labs

Hypothesis

OpenAI, Anthropic, Mistral

Hardware buyer vs API consumer
Potential hardware buyers or capacity contractors at scale; confirm each lab's procurement model rather than infer buying intent from API sales.
Top inference workloads
Production serving of flagship models, reasoning traces, agentic workloads.
GTM route
Executive-level strategic partnership with deep kernel/compiler engineering engagement.
Qualification questions
Which architectures are frozen for 12–24 months? Who owns the serving stack?
Proof of ROI needed
Cost per output token and tokens/sec/user on their own models, in their stack.
Risks
Architecture churn; in-house silicon programs; very high technical bar.

Hyperscalers

Hypothesis

AWS, Microsoft Azure, Google Cloud

Hardware buyer vs API consumer
Possible buyers, but each has in-house accelerators (Trainium/Inferentia, Maia, TPU) that compete.
Top inference workloads
Marketplace or partner-hosted inference offerings.
GTM route
Partner/marketplace and sovereign or regulated-cloud channels rather than core fleet displacement.
Qualification questions
Is there a gap their in-house silicon does not serve? Marketplace economics?
Proof of ROI needed
Customer demand pull plus differentiated SKU margin.
Risks
Direct competitor dynamics; long certification cycles.

Inference & AI platforms

Hypothesis

Databricks, Snowflake, Perplexity

Hardware buyer vs API consumer
Often API consumers or cloud renters today; a subset may buy or reserve dedicated capacity.
Top inference workloads
RAG, retrieval-augmented search answers, data-platform agents, embedded LLM features.
GTM route
Via neocloud/cloud capacity first; direct where volume justifies dedicated racks.
Qualification questions
Are they buying tokens, GPUs-by-the-hour, or hardware? Who pays for capacity?
Proof of ROI needed
Gross-margin improvement on AI features at constant latency.
Risks
May never be hardware buyers; model-provider dependency.

Financial services

Hypothesis

Citadel Securities, Hudson River Trading, Two Sigma

Hardware buyer vs API consumer
Plausible direct buyers with on-prem data centers and strict data control.
Top inference workloads
Research assistants, document/news processing, coding agents, low-latency model serving.
GTM route
Direct technical sale anchored on a published peer reference.
Qualification questions
Inference volume vs. training? Latency vs. throughput priority? Colo power availability?
Proof of ROI needed
Measured throughput and latency on firm-specific models in a paid pilot.
Risks
Secretive buyers; small fleets relative to labs; bespoke requirements.

Weighted account qualification

Illustrative framework for prioritizing prospect hypotheses. It is not an assessment of any real company and never a customer score.

Inference volume · 25%3/5

Sustained token volume determines whether dedicated capacity is economically meaningful.

Workload / model fit · 25%3/5

Purpose-built hardware only matters if the buyer's models and serving patterns are supported.

TCO advantage · 20%3/5

A credible, customer-measured cost or capacity gap is what funds a switch.

Procurement / deployment readiness · 15%3/5

Power, space, budget cycle and an executive owner determine time-to-revenue.

Expansion potential · 15%3/5

Lighthouse accounts matter most when a pilot can become multiyear capacity.

Weighted fit
50
out of 100

Qualify further: close evidence gaps

05 · Government & defense expansion

A longer-cycle, partner-led opportunity.

Hypotheses, not existing sales. Government adoption would most plausibly run through clouds and approved integrators rather than direct hardware sales.

DoD integrators

Booz Allen, Leidos, SAIC

Platform partners

Palantir, Anduril

Defense primes

Lockheed Martin, Northrop Grumman, RTX

DoD enterprise IT

Agency CIO and enterprise AI programs

Civilian government

Agencies with document-heavy mission work

Federal labs

National and service research laboratories

Top candidate use cases

  • Secure retrieval over approved documents and technical knowledge
  • Coding agents for software modernization
  • Maintenance and logistics assistance
  • Human-reviewed cyber investigation support
  • Research and analysis acceleration

Compliance caveat

Hardware itself does not automatically require FedRAMP. A cloud service or hosted environment delivering it may require authorization (FedRAMP, DoD Impact Levels). Security controls, supply-chain provenance and hosting must each be assessed. Indirect routes — authorized cloud providers and cleared integrators — are the realistic early path.

Supporting view

Southeast regional corridor

An illustrative cluster of defense, research and data-center activity worth mapping for partner coverage. Locations are approximate and indicative only.

Norfolk / Hampton RoadsRaleigh-Durham / RTPOak RidgeHuntsvilleAtlantaCharlestonTampa
06 · Inference economics

Model the customer's math — with their numbers.

A working calculator for structuring a TCO conversation. It shows how unit cost, throughput and utilization combine; it proves nothing until the inputs are measured.
NOT ETCHED PRICING OR BENCHMARKS. Every default is an illustrative placeholder chosen to demonstrate the model. Replace with customer-measured data before drawing any conclusion.

Volume & unit cost

Energy (optional)

Capacity

Illustrative output

Existing monthly spend
$1,000,000
Alternative monthly spend
$600,000
Monthly savings
$400,000
Savings %
40%
Annual savings
$4,800,000
Annual output tokens
1,200B
Required peak throughput at chosen utilization
64,300 output tokens/sec at 60% utilization
2
racks today
1
racks proposed
1
difference
  • • Output tokens only; input/prefill tokens, batching, latency SLOs and context length can change real throughput materially.
  • • Spend uses monthly output tokens; rack count uses the separate target capacity.
  • • Excludes migration, software porting, financing, facility, networking and staffing costs.
  • • Lower unit prices can raise demand — total spend may rise even as cost per token falls.
07 · Proof & risk register

What is confirmed, and what still has to be earned.

ClaimSource / dateStatus
First rack shipped to Jane Street by the announcement dateEtched, Aug 18 2026Sourced
$700M raise led by Jane StreetEtched, Aug 18 2026Sourced
>$1B customer contracts (company claim)Press release via HPCwire, Aug 18 2026Sourced
LVI / CSM frontier inference cluster architectureEtched, Jun 30 2026Sourced
Inference volume and compute mix depend on adoption and workloadIndustry context; no quantified market forecast assertedIndustry context
Quantified customer savingsNo independent validation cited here — not assertedHypothesis
Industry context

Manufacturing & supply

Advanced-node wafer, HBM and packaging capacity are contested industry-wide.

Hypothesis

Model compatibility

Supported architectures, precisions, kernels and serving patterns must be validated; current company claims extend beyond transformer-only designs.

Industry context

Competition

NVIDIA and AMD roadmaps plus hyperscaler in-house silicon (TPU, Trainium, Maia).

Industry context

Power & utilization

Economics depend on sustained utilization and available data-center power.

Hypothesis

Customer concentration

Early volume may concentrate in few accounts; one named public reference to date in this review.

Hypothesis

Capex & financing

Buyers finance accelerator fleets; lender familiarity with new silicon may lag GPUs.

08 · 30 / 60 / 90

A GTM plan built on acceptance gates.

Days 1–30

Qualify

  • → Map likely hardware buyers vs API consumers
  • → Confirm supported model/workload envelope with engineering
  • → Score and rank a short list of lighthouse targets
Days 31–60

Prove

  • → Joint technical benchmark plans on customer models
  • → Define success metrics: cost/token, latency, tokens/MW
  • → Executive sponsor and lighthouse plan per account
Days 61–90

Commit

  • → Paid pilots with written customer acceptance gates
  • → Convert passed gates into capacity commitments
  • → Expansion map: sites, power, multiyear roadmap
09 · Why I built this

Chris McCool

26+ years in technology sales, including Netezza purpose-built data-warehouse appliances before and after the IBM acquisition, and IBM Data, Analytics and AI. Complex solution selling at $5M+ in annual sales; helped grow a company through a $40M exit in 2022.

I'm now exploring senior enterprise GTM roles across AI, and built this as an independent work sample of how I'd approach the market.

The Netezza pattern — an analogy, not an equivalence

1
Workload-specific performance
2
Enterprise ROI
3
Technical benchmark
4
Multiyear expansion

Inference silicon differs in scale, supply chain, buyers and capital intensity. What transfers is the sales motion: prove a workload advantage on the customer's own data, translate it into ROI, and earn expansion.

The commercial opportunity is not simply faster silicon. It is measurable intelligence delivered at scale.