Internal · Technology & Operations · July 2026

The AeX technology stack.

A Mac-first, cloud-native operating stack with an AI overlay — engineered to convert market intelligence into three proprietary assets: data, intelligence, and inventory. Priced per person, scored against the mandate, with recommendations and a sequenced build plan.

Team model 5 employees (lean) Data seats 1–2 shared terminals Basis current list / annual-commit pricing, July 2026 Classification pre-decisional
~$2,650
Per person / month
Blended all-in, incl. amortized hardware
~$159K
Annual run-rate
5-person firm, steady state
~73%
Is market data
~$116K of ~$159K
~$49K
One-time hardware
Full fleet capex, Year 1
The one decision that matters. Strip out market data and the entire remaining stack — hardware, AI, warehouse, CRM, security, everything — costs roughly $720 per person per month. Market data alone is ~$1,930 per person per month. This budgeting exercise is really a market-data sequencing exercise; everything else is rounding error by comparison. Sequence the feeds to commercial reality (Recommendations 1–2) and the same firm runs for well under half the headline number until a live trading desk justifies the full data spend.
01 · Read

The architecture is right. The budget is a data decision.

The proposed stack is well-architected and on-thesis: it maps cleanly to a knowledge layer (Obsidian) → enrichment (AI) → warehouse (Snowflake) → surfacing (Hex), on Apple hardware, operable from anywhere. The recurring software and AI layer is inexpensive and best-in-class. The entire economic weight of the stack sits in one place: market data. Keep the design; make the money conversation a sequencing conversation.

02 · Cost model, per person

Four buckets: universal, shared, market data, hardware.

Modeled on 5 employees. Software prices are current list / annual-commit rates. Market-data figures are estimates — those vendors quote privately and bundle by module and licence type; treat them as planning ranges, not quotes.

A. Universal software & AI — one seat per person
ToolTier$ / person / mo$ / person / yr
Google WorkspaceBusiness Plus22.00264
Anthropic ClaudeTeam (annual)25.00300
OpenAI ChatGPTTeam30.00360
PerplexityEnterprise Pro40.00480
ClickUpBusiness12.00144
ObsidianCommercial licence4.1750
1PasswordBusiness8.0096
OktaWorkforce Identity15.00180
CrowdStrikeFalcon Pro (~1.5 endpts)8.33100
Universal subtotal164.501,974
B. Shared platforms — allocated across 5 people
PlatformBasisCompany / mo$ / person / mo
HubSpotSales Hub Pro × 3 seats27054.00
HexTeam × 3 editors + compute37575.00
SnowflakeStandard, light usage500100.00
AWS S3Storage + egress10020.00
QuickBooks OnlinePlus (1 company)11523.00
Banking feesJPM / BofA / Mercury~5010.00
Google DrivePooled w/ Workspace00.00
USGS dataPublic00.00
Shared subtotal1,410282.00
C. Market data — 1–2 shared seats (lean model)
The cost center. Figures are estimates; all six vendors negotiate privately. USGS is free and foundational.
FeedBasisCompany / yr (est.)$ / person / mo
Bloomberg Terminal1 seat @ $31,98031,980533
Benchmark Mineral IntelligenceUp-to-10-user tier~30,000500
S&P Global Commodity InsightsPlatts, 1 sub~25,000417
FastmarketsMetals / battery pkg~18,000300
LME Data ServicesNon-display licence~8,000133
CME Group Data ServicesNon-pro market data~3,00050
USGSPublic reference00
Market-data subtotal~115,9801,933
D. Hardware — capital, then amortized
Fleet: 5 × MacBook Pro, 5 × iPhone, 2 × MacBook Air. Configured-build estimates, amortized straight-line over 36 months.
DeviceConfigUnit (est.)QtyTotal
MacBook Pro 14"M5 Max, 40-GPU, 128GB, 8TB~7,200536,000
iPhone 17 Pro1TB~1,49957,495
MacBook Air 13"M5, 32GB, 4TB †~2,70025,400
One-time capex~48,895
Amortized (36 mo) → per person / mo~272

† Spec flag: the MacBook Air currently maxes at 2TB storage (and 32GB memory), not 4TB. A 4TB build is only available on the MacBook Pro line. Confirm configured Apple prices at order.

Blended: ~$2,651 / person / month · ~$31,800 / person / year · ~$159K company / year.
Year 1 also carries one-offs outside the run-rate: ~$48,900 hardware capex (unless financed) and ~$1,500 HubSpot onboarding. Excluding market data entirely, the blended figure falls to ~$718 / person / month.
03 · Scorecard

Each component, scored 1–10 on fit to the mandate.

Rated on fit to AeX's three proprietary assets — data, intelligence, inventory — plus the Mac-first and remote-operability requirements. High scores mean on-thesis and well-chosen; lower scores flag weak strategic fit or a better option below.

ComponentScoreRationaleCall
MacBook Pro (M5 Max)9Workstation-class local compute for on-device AI and heavy analysis; anchors the Mac-first model.Keep
MacBook Air (M5)8Ideal travel/remote unit — but correct the 4TB spec (Air maxes at 2TB).Adjust
iPhone 17 Pro8Mobile command of ops and dashboards from anywhere; supports the remote mandate.Keep
Google Workspace / Drive8Collaboration and identity backbone; Business Plus adds Vault/eDiscovery for compliance.Keep
Anthropic (Claude)8Strongest agentic fit for the Obsidian knowledge layer. Core to proprietary intelligence.Keep
OpenAI (ChatGPT)7Capable, but largely redundant with Claude at a universal seat. Consolidate.Adjust
Perplexity7Fast sourced research and market scanning; feeds the data-gathering front end.Keep
AWS S38Durable cloud-native store — the system of record beneath the proprietary-data asset.Keep
Obsidian Vault9The knowledge layer itself: Mac-native, portable, plain-text, AI-legible. Central to the thesis.Keep
Snowflake7Right long-term warehouse; likely heavier than a 5-person firm needs on day one.Adjust
HubSpot6Sound CRM, but Sales Hub Pro is over-specced for a small counterparty book. Start Starter.Adjust
ClickUp7Solid ops/workflow layer; watch overlap with Google Workspace and Obsidian tasks.Keep
Hex8Directly delivers the executive-dashboard goal; notebook-to-app fits an analyst desk.Keep
Banking (JPM / BofA / Mercury)6Necessary, but three institutions is heavy ops overhead at 5 people. Rationalize.Adjust
QuickBooks7Standard, adequate; revisit vs. NetSuite only when trade accounting/consolidation grows.Keep
Bloomberg9Gold standard and broadly overlapping — but $32K/seat. Defer until a live desk justifies it.Sequence
Benchmark Mineral Intelligence10The most on-thesis feed listed — battery/rare-earth supply-chain and price intelligence. Lead here.Keep
Fastmarkets9Deep critical-minerals price assessments; core to proprietary pricing intelligence.Keep
S&P Global Commodity Insights8Broad Platts coverage; strong but overlaps Bloomberg/Fastmarkets — sequence.Sequence
LME Data Services8Authoritative for exchange-traded base metals; licence-tier (display vs. derived) matters.Keep
CME Group Data Services7Useful for futures/hedging; low cost. Add when hedging activity begins.Keep
USGS8Free, foundational reference on production, reserves, and trade. No reason not to.Keep
Okta7Right identity backbone; arguably early at 5 people when Workspace SSO + 1Password covers most.Adjust
1Password9Essential, cheap, Mac-native secrets management. No change.Keep
CrowdStrike8Enterprise-grade endpoint protection consistent with a security-conscious posture; premium but justified.Keep
04 · Architecture read

One spine every tool should serve.

Capture (Perplexity, feeds, Web Clipper, transcripts) → Knowledge layer (Obsidian Vault) → AI enrichment (Claude, ChatGPT) → Store (AWS S3 raw + Snowflake structured) → Surface (Hex dashboards, ClickUp workflows, iPhone) → Act (HubSpot counterparties, trade & inventory decisions).

Two structural observations. First, there is a clean split between unstructured intelligence (Obsidian, S3) and structured data (Snowflake, Hex) — good, and it maps to the proprietary-data vs. proprietary-intelligence assets. Second, the proprietary-inventory asset has no dedicated system in the stack yet: physical position, offtake contracts, warehouse receipts, and reserve accounting have nowhere structured to live. That is the notable gap (Recommendation 10).

05 · Recommendations

Ordered by financial and strategic impact.

Sequence market data to commercial reality. This is 73% of run-rate and the only real lever. Lead with the on-thesis, team-priced feeds — Benchmark + Fastmarkets + USGS — which cover critical-minerals pricing and supply chains for a fraction of the total. Add LME/CME when hedging begins. Defer Bloomberg ($32K/seat) until a live trading desk needs real-time cross-asset depth; it broadly overlaps LME, CME, and Fastmarkets for this use. Deferring Bloomberg + S&P in year one saves ~$57K.

Buy data as shared, licensed seats — and scrutinize the licence tier. Exchanges and price-reporting agencies charge very differently for display vs. non-display / derived use, and redistribution — surfacing a price inside a Hex dashboard others view — can trip a higher tier. Negotiate on intended use before signing; at this size it moves price more than seat count.

Consolidate the AI layer. Three universal AI subscriptions ($95/person/mo) is redundant. Recommended: Claude as the universal agentic + knowledge-layer tool, Perplexity for sourced research, and ChatGPT on just 1–2 power seats. Give 1–2 heavy users Claude Max ($100–200/mo) instead of a universal ChatGPT seat. ~$1,500–2,000/yr saved with no capability loss.

Defer Snowflake; start on a Mac-native warehouse. At 5 people and early data volume, DuckDB / MotherDuck gives a columnar warehouse that runs locally on the M5 machines and in the cloud, at a fraction of Snowflake's cost — and Hex connects to it natively. Graduate to Snowflake when volume, concurrency, or governance demands it. ~$5–6K/yr deferred with no loss of the dashboard layer.

Right-size HubSpot. Sales Hub Pro ($90/seat + $1,500 onboarding) is built for outbound sales teams. A merchant with a concentrated counterparty book can start on Starter (~$20/seat) and upgrade when pipeline complexity — not headcount — demands it. ~$2,500/yr + onboarding saved.

Hold Okta until ~10 people or a compliance trigger. Google Workspace SSO plus 1Password covers identity and secrets for a 5-person firm. Okta earns its keep at scale, on lifecycle/de-provisioning volume, or when a counterparty or regulator requires it. Keep it on the roadmap, not the launch list. ~$900/yr deferred.

Add the missing layer: Mac fleet management (MDM). A remote, security-conscious Mac fleet needs device management — Kandji or Jamf (~$4–9/device/mo) — to enforce disk encryption, push CrowdStrike, and remote-wipe a lost machine. This is the notable omission in an otherwise strong security posture. Pair with Tailscale for zero-config secure remote access.

Formalize backup and continuity. 8TB local drives holding proprietary research need an offsite backup path beyond S3 (e.g. Backblaze + Time Machine). Define it before, not after, the first drive failure.

Rationalize banking. Three institutions at 5 people is operational drag. A common pattern: Mercury for day-to-day operating and cards, one money-center bank (JPM) for trade finance, LCs, and scale as inventory grows. Revisit BofA's role — banking redundancy is deliberate at scale, premature at five people.

Give "proprietary inventory" a home. The third asset has no system. Before physical trading begins you'll need somewhere to hold positions, offtake contracts, warehouse receipts, assay/QA records, and reserve accounting — plus a virtual data room for deal diligence. Options range from a structured Snowflake/DuckDB schema + Hex to a purpose-built commodity trade & risk management (CTRM) tool as volume justifies. Decide the day-one version now.

06 · Implementation checklist

Identity and security before data flows; cheap feeds before expensive ones.

Phase 1 · Foundation (Week 1–2)
Phase 2 · Knowledge & intelligence (Week 2–4)
Phase 3 · Market data — sequenced (Week 3–6)
Phase 4 · Operations & commercial (Week 4–8)
Phase 5 · Governance (ongoing)
Bottom line. Keep the architecture — it is genuinely well-designed and on-thesis. Correct the MacBook Air spec, close the MDM and inventory-system gaps, consolidate the AI seats, and treat the entire budget as a market-data sequencing decision. Do that and AeX runs a globally credible, Mac-first, remotely operable intelligence stack for roughly $70–90K/year until a live trading desk justifies stepping up to the full ~$159K data-heavy configuration.

Sources & method

  • Software pricing (current list / annual-commit, July 2026): Bloomberg, Hex, HubSpot, Snowflake, CrowdStrike, Okta, 1Password, Google Workspace, QuickBooks, ChatGPT, Claude, Perplexity, ClickUp, Apple. Bloomberg · Hex · HubSpot · Snowflake · Okta
  • Market-data figures (Benchmark, Fastmarkets, S&P Global Commodity Insights, LME, CME): directional estimates — these vendors quote privately and bundle by module and licence type. Treat as planning ranges, not quotes. Benchmark plans
  • Hardware: configured-build estimates to be confirmed at order; note the MacBook Air storage-cap correction. MacBook Pro M5
  • Model basis: 5-person team, lean 1–2 shared data-seat assumption; hardware amortized straight-line over 36 months. All figures reconcile to a ~$159K annual run-rate and ~$2,651 blended per-person-per-month.

Directional planning tool, not a procurement quote. Market-data costs in particular are estimates pending vendor negotiation. Prepared July 2026.

AeX internal · Technology stack brief · July 2026 · not legal, tax, or investment advice