Signal dashboard
Loading...
📊 Signal leaderboard
Loading signals…
🔍 Portfolio stack scan
Click "⚡ Run scan" to scan companies across 24 investor portfolios for tech stack signals, job boards, and GitHub presence.
Models
OpenRouter market · startup adoption · pricing signals
Total models
—
on OpenRouter
Providers
—
unique
Cheapest prompt
—
per 1M tokens
Free tier models
—
available free
Startup model adoption
Run a scan to see which models startups are actually using
Inference cost snapshot
Loading...
Provider ranking
Loading...
Pricing signals
Loading...
All tracked models
Loading...
Hardware & infrastructure
GPU rental · cloud compute · memory pricing · live & manual signals
Silicon iGPU contract and spot rental pricing. 1yr contracts from SemiAnalysis public index. Neocloud spot vs hyperscaler on-demand spread signals demand tightness.
GPU rental pricing
| GPU | Contract | Price/GPU/hr iRental cost per GPU per hour on a 1-year contract. Compare to on-demand hyperscaler pricing to understand the neocloud vs hyperscaler spread. | vs Oct 2025 | Source |
|---|---|---|---|---|
| H100 80GB | 1-year | $2.35/hr | ↑ +38% from $1.70 | SemiAnalysis · Mar 2026 |
| H100 80GB | On-demand | $12.29/hr | vs neocloud | AWS p5 · us-east-1 |
Full GPU index (H200, B200, B300, GB200, MI300) available via SemiAnalysis institutional subscription.
View public dashboard →
Cloud compute iOn-demand GPU instance pricing from AWS, GCP, and Azure. Compare hyperscaler pricing to understand who is winning AI workload market share.
GPU instances
Fetching cloud pricing...
CPU compute iGeneral purpose CPU instances. Intel vs AMD vs Arm (Graviton/Ampere/Cobalt) pricing signals which architecture is winning on price-performance. Arm is typically 20-30% cheaper than Intel for equivalent cores.
CPU instances
Loading...
Memory iHBM and DRAM pricing. HBM (High Bandwidth Memory) is the real AI bottleneck — it's the memory inside H100/H200 GPUs. When HBM supply tightens, GPU supply tightens. Track Samsung, SK Hynix, Micron via TrendForce.
Memory pricing signals
| Memory type iHBM = High Bandwidth Memory inside AI GPUs (H100, B200). DDR5 server = standard server DRAM. Both are constrained by the same fab capacity at Samsung, SK Hynix, Micron. | Signal / price | Trend | Source & date |
|---|---|---|---|
| HBM3e (per chip) | ~$300 | ↑ vs HBM3 ~$150 | NAND Research · Nov 2025 |
| HBM4 (per chip, est.) | ~$500 | ↑ +67% vs HBM3e | NAND Research · Nov 2025 |
| HBM3e vs DDR5 premium | 4–5x DDR5 | → narrowing to 1–2x by end 2026 | TrendForce · Dec 2025 |
| Server DDR5 contract | +18–23% QoQ | ↑ above forecast | TrendForce · Q4 2025 |
| DDR5 spot (16Gb eTT) | ~$20–23.50 | ↑ quadrupled since Sep 2025 | DRAMeXchange · Q1 2026 |
| DDR4 (16Gb eTT) | ~$13–14.80 | ↑ +10%/wk Oct 2025 | DRAMeXchange · Oct 2025 |
Exact contract prices require TrendForce Gold+ subscription. Data above from public press releases and reporting.
Inference cost iCost per million tokens by provider. Inference commoditization is the key trend — when this drops it signals the GPU stack is becoming more efficient and competitive.
OpenRouter inference pricing
Loading...
Hiring signals
AI role trends · job board platforms · skills in demand · from last scan
Run a scan to populate hiring signals across all portfolio companies
Business stack
Payments · analytics · CRM · comms · RevOps · from last scan
Run a scan to populate business stack signals across all portfolio companies
Startup cohort
Investor universe tracked by STAX
Weekly brief
Your STAX signal summary — ready to publish
Loading brief…
How STAX works
Signal detection methodology · what we read, what we infer, and what we miss
⚡
Signals ≠ Adoption
STAX does not measure technology adoption. It measures what's publicly observable. These signals are imperfect, incomplete, and sometimes wrong. A signal tells you a technology left a footprint somewhere we could see — it doesn't tell you how central that technology is to the business, whether it's in production, or whether the company even chose it intentionally. Use them as directional indicators, not ground truth.
—
companies tracked
—
investor portfolios
8
detection passes
The core idea
STAX reads what's leaking out, not what's going on inside.
Everything you see here is inferred from publicly visible data — what companies expose on their websites, job boards, and GitHub profiles. That's a feature, not a bug: public signals are often the earliest indicators of a technology spreading through a market, before any analyst report acknowledges it.
Venture-backed startups move fast and signal early. The infrastructure choices a company makes at seed and Series A often reflect where the market is heading 18–24 months before it becomes consensus. STAX is designed to surface those choices at scale, across portfolios, before they show up in analyst reports.
Detection pipeline
Domain
→
HTTP headers
→
DNS / MX
→
IP ranges
→
Subdomains
→
Job boards
→
GitHub
→
AI libs
→
Biz stack
→
Signals
🌐
HTTP headers
Response headers, cookies, and page metadata reveal CDNs, hosting providers, analytics platforms, and frameworks without any self-reporting.
CF-Ray → Cloudflare · X-Powered-By → framework
🔍
DNS & MX records
TXT records reveal ownership verification (Salesforce, Notion, Atlassian). MX records expose email provider — Google Workspace vs Microsoft 365 vs custom.
📡
IP range matching
We fetch official AWS, GCP, and Cloudflare IP range files and match resolved IPs. Definitive cloud provider detection, even through custom domains.
🔗
Subdomain scanning
Common subdomains (app., api., docs., status.) often expose different infrastructure than the marketing site, revealing the actual product stack.
💼
Job board detection
We detect which ATS a company uses (Greenhouse, Lever, Ashby, Workday) and parse job postings for AI keywords: RAG, fine-tuning, LLM, inference, DPO, vector database. Companies hire for what they're building, not what they've shipped.
🐙
GitHub deep scan
Domain-matched GitHub orgs. We parse dependency files (package.json, requirements.txt, Gemfile) in top repos to detect AI libraries and infrastructure packages actually in use.
🧠
AI & open-weights detection
Page content and job descriptions scanned for AI framework references — LangChain, LlamaIndex, OpenAI, Anthropic, and open-weights models: Llama, Qwen, Gemma, Ollama, vLLM, self-hosted LLM.
🏢
Business stack signals
Beyond infrastructure: payments (Stripe, Braintree, Adyen), auth (Auth0, WorkOS), CRM (Salesforce, HubSpot), observability (Datadog, New Relic, Grafana), comms (Intercom, Zendesk), and more.
What we detect
How to read the leaderboard
The Signal Leaderboard shows what percentage of scanned companies expose each technology signal. A technology at 25% doesn't mean 25% of startups "use" it — it means 25% of their public-facing presence reveals that technology. Think of it as market signal density, not market share.
The comparison view (VS [date]) shows week-over-week change. A technology moving from 14% to 18% in one scan cycle is a meaningful leading indicator — especially if it's concentrated in a specific investor's portfolio or cohort. Delta badges highlight meaningful movement.
What these signals don't tell you
⚡
Signal ≠ usage
A detected signal means we found a footprint — not that the technology is central to the business. A Stripe script on a marketing page is not the same as Stripe powering the product.
🔇
Absence ≠ non-adoption
Many enterprise and stealth companies deliberately minimize their public footprint. High signal density is meaningful; low signal density is inconclusive.
🟠
Cloudflare creates blind spots
Companies behind Cloudflare's proxy suppress headers that would otherwise reveal hosting and framework. High Cloudflare adoption is itself meaningful, but it also hides other signals.
🏠
Marketing site ≠ product infra
A company might run AWS internally but serve their marketing site on Vercel. We detect what they show the world, not what powers their core product.
🌫️
Job titles are fuzzy
"AI Engineer" means different things at different companies. We detect keyword presence, not intent or seniority.
🐙
GitHub matching is domain-based
Companies with non-obvious org names or multiple repos under different names may not be detected correctly.
💻
GPU pricing is directional only
Hyperscaler prices bundle egress, networking, and SLAs. Neocloud prices are typically raw compute only. Use as directional signal, not procurement guidance.
🗺️
Coverage is US-biased
We only track companies whose investors publish portfolios publicly. The dataset skews toward US-based, early-stage, VC-backed startups.
Coverage
STAX currently tracks 6,000+ companies across 60+ investor portfolios including YC, a16z, Sequoia, Greylock, Lerer Hippeau, Bessemer, Lightspeed, General Catalyst, Index, Flybridge, ERA, Eniac, Factorial, Betaworks, NEA, Battery, Benchmark, Kleiner Perkins, First Round, Lux Capital, Madrona, BoxGroup, AlleyCorp, and many others. New portfolios are added regularly.
We're building this in public. If you have feedback on methodology, signal quality, or portfolios you'd like to see added, reach out at From the Porch.
Built by Porch Capital · porch.capital
Trends
How signals move across the portfolio over time
Category
Tools (up to 5)
Range
View
Select tools above to chart adoption over time
Each point is a full portfolio scan (≥3,000 companies). % = share of scanned companies with the signal detected. Tools are tracked per category; gaps indicate zero detections on that date.
Inbox
Import signals
Manually add GPU pricing, stack data, and other signals
Add signal
Category
Name / provider
Value / metric
Notes (optional)
Imported signals
No signals imported yet