A synthetic operating model for translating Vercel usage units into product-level unit economics, owner accountability, and launch gates.
The artifact demonstrates how I would structure the system, not what Vercel's margins are.
Thesis
The harder problem is building a cost accountability layer that lets EPD see margin consequences before usage scales: requests, active CPU, provisioned memory, transfer, Blob operations, AI Gateway tokens, cache misses, provider mix, and customer contract terms.
A useful finance partner should separate metered customer usage from internal cost drivers, then turn that map into pricing, forecasting, roadmap, and launch decisions.
Important distinction
AI Gateway should be tracked differently from v0-style generation economics. Gateway pricing is zero-markup/pay-as-you-go in public docs, including BYOK. The finance question is pass-through accuracy, provider mix, payment leakage, and usage forecast. v0-style workloads are product COGS: cost per accepted output, retained user, and generation.
Why this matters
Every invoice line should resolve to a product, workload, customer segment, and EPD owner. If not, it belongs in an explicit quarantine bucket with an aging clock. Unallocated cost is not overhead; it is a broken instrumentation workflow.
Model AI Gateway credits, BYOK, provider list price, and v0-style generation as separate economics.
Cost metrics should point to an action: cache, route, reduce active CPU, resize memory, move workload, reprice.
Direct tag beats allocation formula. Formula beats model. Model beats unknown. Unknown gets aged and escalated.
Included usage, overages, enterprise exceptions, and AI credits must flow back into unit economics.
First-week EPD questions
Source tables
This is the smallest schema I would want before calling a cost view “EPD-ready.”
event_ts
team_id
project_id
product_surface
workload_type
region
edge_requests
active_cpu_ms
provisioned_memory_gb_ms
transfer_gb
blob_simple_ops
input_tokens
output_tokens
provider
model_id
cache_status
invoice_month
vendor
sku
region
usage_unit
usage_qty
cost_usd
direct_tag_key
allocation_method
confidence_score
product_surface
workload_type
epd_owner
finance_owner
pricing_owner
margin_bearing_flag
pass_through_flag
launch_gate_required
customer_segment
plan
included_usage_unit
included_qty
overage_rate
ai_credit_terms
enterprise_exception
Allocation ladder
case
when direct_tag_key is not null
then 'direct'
when product_surface and workload_type are present
then 'deterministic_usage_map'
when only segment / region can be inferred
then 'modeled_allocation'
else 'quarantine_unallocated'
end as allocation_methodAI Gateway
Gateway needs provider/model mix, token volume, BYOK handling, and payment leakage tracked separately from v0-style product COGS.
EPD levers
Unit economics map
| Surface | Finance unit | Business question | Do not blend with |
|---|
Launch review
The gate is not meant to slow EPD down. It is meant to make the margin shape visible early enough that Engineering, Product, Pricing, and Finance can choose the right tradeoff deliberately.
Edge request, active CPU, GB transfer, Blob op, token, generation, or customer-specific contract unit.
Product, engineering owner, finance owner, and pricing owner are assigned before scale-up.
AI Gateway credits/BYOK are separated from margin-bearing AI product COGS.
Blob, CDN, and origin-transfer costs show hit/miss behavior and file-size edge cases.
Unallocated spend must be below 5% or have a named remediation owner and date.
Why this artifact exists
This is the kind of artifact I would rather send than a generic cover letter: a small example of how I think about Vercel's EPD finance problem, grounded in public product mechanics and honest about what I cannot know from the outside.