Self-Hosted AI Deployment in San Francisco, California

Cybersecurity / SaaS / Self-Hosted AI / San Francisco

A San Francisco-based cybersecurity company (USD $82M ARR, providing cloud security posture management — CSPM — and cloud workload protection, 480 employees, protecting 2,400 enterprise customers' cloud environments across AWS, Azure, and GCP, Series C with USD $180M raised at USD $1.2B valuation) needed to deploy AI capabilities to transform threat detection, vulnerability prioritisation, and security incident response. The company's fundamental challenge was existential: as a cybersecurity company, sending customer security data to external AI providers would create the exact vulnerability their product was designed to prevent. Core challenges: (1) Security data sovereignty — non-negotiable. The company processed the most sensitive data imaginable: cloud infrastructure configurations (network topology, security group rules, IAM policies, encryption settings — a complete blueprint for attacking each customer's cloud environment), vulnerability scan results (every known vulnerability in every customer's cloud environment — essentially a target list for attackers), security event logs (authentication attempts, API calls, network flows, application logs — detailed records of every action in the customer's cloud), and compliance posture (SOC 2, PCI DSS, HIPAA, ISO 27001 gaps — knowledge of exactly where each customer's security was weakest). If any of this data was exposed through an external AI provider breach, the cybersecurity company would: lose every customer immediately (trust is the product — any data exposure destroys it permanently), face massive litigation (customer security data exposure enabling attacks), and likely cease to exist (a cybersecurity company that caused security breaches has no future). The company's CISO had a simple policy: "no customer security data leaves our infrastructure, ever, for any purpose, including AI processing." (2) Threat detection AI — moving beyond rules. The company's existing threat detection used rule-based policies and statistical anomaly detection — effective for known attack patterns but missing: novel attack techniques (new attack vectors that did not match existing rules), slow-and-low attacks (sophisticated attackers making small changes over weeks or months — individually normal, collectively comprising an attack), misconfiguration chains (individual misconfigurations that were low-risk alone but together created critical attack paths), and context-dependent risks (configurations that were normal for development environments but dangerous in production — requiring contextual understanding). LLM-powered threat detection could analyse security configurations and events with contextual understanding — recognising that a specific IAM policy change was routine maintenance OR the first step of a privilege escalation attack depending on who made the change, when, and what other changes accompanied it. (3) Vulnerability prioritisation — the signal-to-noise problem. The company's customers collectively had 2.8 million active vulnerabilities across their cloud environments. Traditional vulnerability scoring (CVSS) ranked vulnerabilities by inherent severity but not by exploitability in each customer's specific environment. A critical CVSS vulnerability in an isolated, unexposed system was less urgent than a medium CVSS vulnerability in a publicly exposed system with sensitive data. Security teams suffered from "vulnerability fatigue" — overwhelmed by thousands of findings, unable to determine which to fix first. The company needed AI that could: analyse each vulnerability in the context of the customer's specific infrastructure (network exposure, data sensitivity, compensating controls), predict exploit probability (based on threat intelligence, exploit availability, and attacker targeting patterns), and recommend prioritised remediation (the top 20 vulnerabilities to fix today, with automated remediation where possible). This contextualisation required processing the customer's complete cloud configuration — the most sensitive data in the product. (4) Security copilot — analyst augmentation. The company's enterprise customers had security operations teams (SOC) that used the product daily. These analysts needed: natural language querying (asking "which of our production databases are accessible from the internet?" rather than writing complex queries), incident investigation assistance (given a security alert, AI explaining what happened, what the impact could be, and what to investigate next), compliance gap analysis (AI explaining why specific configurations failed compliance checks and how to remediate), and security architecture review (AI analysing proposed infrastructure changes for security implications before deployment). Building a security copilot required deploying an LLM that understood cloud security at expert level — not a general-purpose model making security-sounding statements, but a model that genuinely understood AWS IAM policy syntax, Azure networking, GCP security architecture, and the interplay between them. (5) Pre-IPO requirements — at USD $82M ARR and $1.2B valuation, the company was on a trajectory toward IPO within 24-36 months. AI infrastructure decisions needed to be IPO-ready: SOX-compliant data processing controls, auditable AI governance, clear documentation of how customer data was processed by AI, and data processing agreements with customers that explicitly addressed AI. Self-hosted AI simplified these requirements — "we process your data on our infrastructure using our AI models with no third-party access" was a simple, defensible statement for enterprise security assessments, SOC 2 audits, and S-1 risk factor disclosure.

We deployed a self-hosted AI platform over 10 weeks — purpose-built for cybersecurity operations with the absolute data isolation the company required. (1) Air-gapped AI infrastructure. Hardware: 32 NVIDIA H100 80GB GPUs across 4 server nodes, deployed in the company's primary data centre in Equinix SV5 (Santa Clara — the company's existing production infrastructure). Co-located with the company's security data processing pipeline — AI inference occurring within the same security boundary as all other data processing, with no additional data movement. Configuration: 8x H100 per node, 1.5TB system RAM, 24TB NVMe storage, InfiniBand HDR200 interconnect. Network: air-gapped from public internet. Connected only to the company's internal processing network. No API endpoint accessible from outside the company's infrastructure — the AI was an internal service, not an externally accessible system. Even the company's own employees could not access the AI directly — all interaction through the product's existing security boundary (authenticated, authorised, logged). Security: the AI infrastructure meeting the same security standards as the company's production security data processing — SOC 2 Type II controls, encryption at rest and in transit, access controls, and comprehensive audit logging. (2) Contextual threat detection AI. Attack path analysis: Llama 3 405B fine-tuned on the company's security assessment database (4.2 million cloud configurations from 2,400 customer environments, annotated with security assessment outcomes). The model learning to identify: multi-step attack paths (a publicly exposed EC2 instance with an unpatched vulnerability, connected to an internal database via overly permissive security group, with cross-account IAM role allowing privilege escalation — each element individually flaggable but the combination representing a critical attack chain that rule-based detection missed), identity-based risks (IAM policies that granted excessive permissions through complex policy combinations — policy A granting S3 access, policy B granting KMS access, combined enabling decryption of all customer data), and temporal patterns (configurations that changed gradually toward insecure states — security group rules accumulating over months, each addition individually approved but collectively creating exposure). Attack path model processing: analysing each customer's cloud environment nightly (complete configuration scan) and identifying critical attack paths ranked by exploitability and potential impact. Novel attack detection: the model identifying 340 previously unknown attack patterns in the first 6 months — configuration combinations that the company's existing rule set did not cover but that the AI recognised as dangerous based on patterns learned from the broader dataset. These discoveries became new detection rules, improving the product for all customers. Misconfiguration context: the AI distinguishing between: configurations that were genuinely dangerous (immediate remediation required), configurations that were acceptable in context (development environment settings that would be dangerous in production but were appropriate for their current use), and configurations that were technically non-compliant but compensated by other controls (a public S3 bucket that was protected by CloudFront with WAF and signed URLs — technically "public" but effectively protected). This contextual understanding reducing false positive alerts by 42 percent — security teams receiving fewer but more actionable findings. (3) Intelligent vulnerability prioritisation. Context-aware scoring: the AI combining CVSS base scores with customer-specific context to produce exploitability-adjusted priority scores: network exposure (was the vulnerable system reachable from the internet, from other VPCs, or only locally?), data sensitivity (what data could an attacker access through this vulnerability — classified using the company's data classification model), compensating controls (were there WAF rules, network ACLs, or IAM restrictions that reduced exploitability?), threat intelligence (was there active exploitation of this vulnerability in the wild — correlated with the company's threat intelligence feeds), and blast radius (if exploited, how many other systems and data stores could be reached from the compromised system?). Prioritisation impact: customers' security teams now fixing the most impactful 5 percent of vulnerabilities first — reducing actual security risk by 68 percent while only remediating 140 vulnerabilities (versus unfocused approach of remediating hundreds with uncertain impact). Automated remediation: for low-risk, well-understood remediations (specific security group rule removals, S3 bucket policy corrections, encryption enablement), the AI generating remediation code (CloudFormation, Terraform, AWS CLI) that security teams could review and execute — or auto-apply for customer-approved remediation categories. 34 percent of vulnerabilities auto-remediable (with customer approval and guardrails). (4) Security copilot: natural language security analysis. Model: Llama 3 70B fine-tuned on: the company's security assessment methodology (how the company's security researchers evaluated cloud configurations), 28,000 customer support interactions (how security analysts asked questions and what answers they needed), cloud provider documentation (AWS, Azure, GCP security documentation — enabling the model to explain configurations with reference to provider-specific concepts), and compliance frameworks (SOC 2, PCI DSS, HIPAA, ISO 27001, CIS Benchmarks — enabling the model to explain compliance gaps with specific framework references). Natural language querying: security analysts could ask questions in plain English: "Which of our S3 buckets contain PII and are accessible from outside the VPC?" — the AI querying the customer's configuration data and returning specific, actionable results. "What changed in our production environment last week that could affect our PCI compliance?" — the AI correlating configuration changes with PCI DSS requirements and highlighting relevant changes. "Explain why this IAM role is flagged as risky" — the AI providing detailed analysis of the specific policy statements, their effective permissions, and the attack scenarios they enabled. Incident investigation: when a security alert fired, the AI providing: incident summary (what happened, based on the alert and surrounding events), blast radius assessment (what systems and data were potentially affected), investigation guidance (what logs to check, what configurations to verify, what indicators of compromise to look for), and remediation recommendation (specific steps to contain and resolve the incident). Security architect review: before deploying infrastructure changes, teams could ask the AI to review Terraform or CloudFormation templates: "Review this Terraform plan for security issues" — the AI identifying misconfigurations, missing encryption, overly permissive access, and compliance violations before deployment. (5) SOX and IPO preparation. AI governance: comprehensive documentation of AI system architecture, data flows, model governance, and security controls — designed for: SOC 2 audit (the company's SOC 2 Type II audit covering AI systems — auditors able to examine AI data handling, access controls, and processing documentation), customer security assessments (enterprise customers conducting vendor security assessments — clear documentation showing that customer data was processed only on the company's infrastructure with no third-party AI provider involvement), S-1 risk factor disclosure (preparing for IPO, the company's legal team reviewing AI risk factors — self-hosted deployment significantly simplifying risk disclosure compared to third-party AI dependency), and SOX preparation (financial AI applications — revenue forecasting, usage analytics — meeting SOX auditability requirements through self-hosted deployment with comprehensive audit trails).

Results over 12 months post-deployment. Product differentiation: the self-hosted AI security copilot became the company's primary competitive differentiator — cited in 72 percent of enterprise deals as a key evaluation factor. Competitive win rate improved from 34 percent to 52 percent against primary competitors (who used cloud AI APIs with customer data sovereignty concerns). Enterprise deal size increased 38 percent (AI capabilities enabling upsell to enterprise tier). Threat detection: novel attack patterns identified: 340 in first 12 months (becoming new detection rules for all customers). False positive rate reduction: 42 percent (customers receiving fewer, more actionable alerts). Customer mean time to remediate critical findings: from 18 days to 4 days (prioritised remediation focusing effort on highest-impact issues). Vulnerability prioritisation: 68 percent risk reduction from remediating top 5 percent of prioritised vulnerabilities. Automated remediation: 34 percent of findings auto-remediable. Customer security posture scores (the company's proprietary metric): improved 45 percent on average across enterprise customers. Security copilot: copilot adoption: 86 percent of enterprise customer security teams using the copilot weekly. Average queries per analyst per day: 12 (indicating deep integration into security workflow). Analyst productivity: 2.4x improvement in findings investigated per analyst per day. Customer NPS: improved from 42 to 68 (copilot cited as primary driver). Revenue impact: ARR growth: from USD $82M to USD $142M (73 percent growth in 12 months). Net revenue retention: from 118 percent to 134 percent (AI features driving expansion). New enterprise logos: 380 (from 220 in prior year — 73 percent increase). AI-attributed revenue: USD $34M ARR (24 percent of total — AI features enabling enterprise tier pricing). Performance: H100 cluster handling nightly configuration analysis for 2,400 customer environments plus real-time copilot queries. Configuration analysis: processing complete cloud environment scan in 45 seconds average per customer. Copilot query latency: median 2.8 seconds for analytical queries, 8.4 seconds for complex attack path analysis. System uptime: 99.99 percent (the company's SLA for security products). Cost: infrastructure investment: USD $3.6M (32 H100 GPUs, servers, networking, storage). Annual operating cost: USD $1.1M. Revenue impact: USD $60M ARR growth (from $82M to $142M) with AI as primary growth driver. The self-hosted investment was immaterial relative to revenue growth — the CFO describing it as "the highest-ROI infrastructure investment in company history." IPO preparation: SOC 2 Type II audit covering AI systems — zero findings. Customer security assessments: simplified data processing narrative ("all processing on our infrastructure, no third-party AI access") accelerating enterprise procurement cycles by average 3 weeks. S-1 AI risk factor section: drafted with clear, defensible language enabled by self-hosted architecture.