Reference stacks

Complete AI and hardware stack recommendations by team type.

Reviews answer whether a product is good. Stack pages answer whether the combination is sensible for the team, budget, data risk, and operating burden.

Reference stacks

Start from the workflow, then choose the model, app, hardware, and controls.

One technical founder shipping product, demos, support, and research without a platform team.

Solo AI founder stack

Use hosted reasoning only for high-stakes work, keep fast models for drafts, and do not buy a rack before workloads repeat.

Lean monthly software spend with an optional 24GB workstation once local evals happen weekly.

Primary model: Frontier reasoning model Escalate code review, architecture, customer-impacting analysis, and research synthesis.
Daily model: Fast utility model Handle summaries, drafts, extraction, labels, and support macros cheaply.
Memory layer: AI notebook app Keep founder calls and research searchable without pretending it is company policy.
Hardware: Thin 14-inch creator laptop Optimize for travel, calls, demos, and short local tests.

Do not make the slowest model the default for every prompt.
Do not buy server hardware until utilization is predictable.
Do not let personal meeting memory become the official company source of truth.

Start with model routing and notebook capture; add local GPU only when privacy or repeated eval volume justifies it.

Product engineering team using AI inside pull requests, migrations, bug triage, and release checks.

Engineering review stack

The win is not more generated code; it is better first-pass review with ownership and test evidence intact.

Pay for reasoning on risky diffs and use automation tooling to keep review loops auditable.

Review gate: Frontier reasoning model Use for high-risk diffs, multi-file migrations, architecture changes, and agent recovery.
Workflow layer: Prompt automation toolkit Codify review prompts, evals, handoffs, and approval states.
Fast path: Fast utility model Generate summaries, changelog drafts, labels, and low-risk comments.
Audit: Repository-aware review bot Keep inspected files, test evidence, and reviewer ownership visible.

Do not merge AI changes outside normal code review.
Do not trust a review tool that cannot show what it inspected.
Do not optimize for comment volume; optimize for missed-regression reduction.

Pilot on risky pull requests first, then expand to summaries and triage after false positives are understood.

Team with privacy-sensitive workloads, repeated local evals, and someone accountable for power, cooling, and service.

Private inference lab stack

Owned hardware is a control decision first and a cost decision second; serviceability beats small benchmark wins.

Capex only makes sense when utilization, power, spares, and operator time are part of the calculation.

Dev loop: 24GB local inference workstation Use for private prompts, quantized model tests, and repeatable evals.
Rack path: 4U rack inference node Use when density, remote management, and service access matter.
Cloud fallback: Hosted fast utility model Keep burst traffic and overflow off the local box.
Operations: Power and cooling plan Treat airflow, noise, rack depth, and remote access as first-class requirements.

Do not place rack-class GPU servers near desks.
Do not compare hardware without including idle time.
Do not skip remote management just because the first node is nearby.

Buy the workstation first for eval loops; move to rack hardware only after utilization and operator ownership are proven.

Agency or implementation team producing client reports, automations, research packs, and recurring workflows.

Agency delivery stack

Reusable prompts and review states matter more than the fanciest model for most agency work.

Keep premium model use behind review gates; spend the savings on workflow templates and client-safe exports.

Workflow layer: Prompt automation toolkit Template repeated deliverables, reviews, and approvals.
Research memory: AI notebook app Capture client context without mixing it into shared policy by default.
Escalation model: Frontier reasoning model Use for difficult synthesis, technical review, and final risk checks.
Client handoff: Governed knowledge base Promote only reviewed outputs into official client materials.

Do not let every client create a new ad hoc workflow.
Do not store sensitive notes without export and deletion clarity.
Do not sell AI speed while hiding review accountability.

Standardize three recurring workflows before adding more model providers or custom apps.

Comparison workbench

Choose two reviews and inspect the tradeoff.

Option AOption B

Decision point	Frontier reasoning model for code and research	24GB local inference workstation	Read this as
Score	9.3	8.8	Frontier reasoning model for code and research
Best for	Senior-code workflows, planning agents, technical research, and review gates.	Model tinkering, privacy-sensitive prototypes, eval runs, and developer labs.	Depends on workload
Buy when	Route hard code review, architecture, agent planning, and research synthesis here first.	Buy when privacy, iteration speed, and repeated local experiments matter every week.	Match to operating constraint
Skip when	Do not use it as the blanket default for extraction, summaries, simple support replies, or bulk drafts.	Skip if you mainly need production concurrency, burst capacity, or models above the card's memory ceiling.	Avoid wrong default
Retest trigger	Retest after major context, tool-use, or price updates.	Retest after driver updates, new 24GB cards, or street-price movement.	Review before purchase

Price and update watch

Signals that change the answer

All signals

GPUs | Buy

24GB GPU street price drops below two months of projected cloud spend

Buy the local workstation only if weekly utilization is already visible.

Recheck used/new warranty, driver stability, PSU headroom, and resale risk before purchase.

AI Models | Retest routing

Frontier model price or latency changes materially

Update escalation rules before changing the default model.

Run code review, research synthesis, support reply, and extraction prompts through the same scorecard.

Laptops | Wait for retest

Laptop BIOS update claims better fan curve or AI performance

Delay a fleet buy until sustained load, battery, and noise are remeasured.

Retest compile loop, video call battery drain, local inference burst, screen behavior, and port fit.