The AI Infrastructure Volcano: What Law Firms Miss While Chasing Speed
While drafting this piece, a Cloudflare outage froze the AI tools I use for background work. Everything stopped for hours. If that had hit an hour before a filing, it wouldn't be a funny story - it would be a professional problem.
Legal is debating hallucinations, privacy, and adoption curves. That's fine. But there's a more basic risk under the surface: AI development can stall, and systems can fail for reasons no lawyer controls. That risk is real, and it's growing.
Where We Are Right Now
The market is selling magic. Vendors promise bigger, smarter models that do more mysterious tasks. Commentators push law firms to adopt fast or fall behind. Money is flooding in, valuations are ballooning, and firms are swapping core workflows for systems they don't own and can't fix.
The assumption is simple: models will get better and work on demand. But there's a hard ceiling that doesn't care about marketing - chips, data centers, and electricity. You can't run AI without them.
The Numbers Don't Support Complacency
Business Insider recently flagged an infrastructure crunch hiding in plain sight. Developers proposed 26 new data center projects that would need an extra 711 gigawatts of power beyond current capacity. That's nearly the same as total summer peak demand across the continental U.S.
If already permitted data centers came online, demand could hit 293 terawatt hours - roughly the power Florida uses in a year. There isn't enough energy capacity to run the silicon required for the AI everyone says is coming. And you don't speed-run utility approvals. As anyone who has dealt with state public service commissions knows, the process can make watching paint dry feel fast.
Even if the grid and data centers expand, that buildout takes years. If AI plateaus while we're building, we end up with too much capacity and not enough demand. If AI keeps growing faster than infrastructure, we get throttling, outages, and higher costs.
What Happens When Demand Outruns Supply
- Slower responses and degraded service during peak times. Think digital traffic jams.
- Throttling and temporary shutdowns when capacity is hit - the tech version of brownouts.
- Higher inference costs as compute becomes scarce, squeezing margins on alternative fee matters.
- Stalled model training, slower product roadmaps, and a plateau in capabilities.
This isn't theoretical. Outages are already happening. Multiply that by a systemic grid or chip shortage and the pain spreads fast.
Why This Hits Legal Hard
Legal work runs on deadlines and client expectations. You can't tell a judge the AI was down. You can't miss discovery because a vendor throttled compute. Clients won't accept "the tool crashed" as a reason for shoddy work.
Many firms are building pricing models around AI efficiency. If models slow or costs spike, the math breaks. Malpractice carriers won't care that your system timed out. Neither will opposing counsel.
Signals From The Market
Even big tech is starting to tap the brakes. As reported by Reuters, Google's CEO warned that if the AI boom stalls, no company is spared. There's some irrationality in the market right now, and the infrastructure strain is a big reason why.
Read the latest from Reuters on AI market risks.
There are smart responses emerging. Some teams are going smaller and closer to the load: modular, decentralized data centers that use existing and underused energy sources like combined heat and power or geothermal. Think local capacity that spins up fast and bridges peak demand. That approach won't solve everything, but it adds resilience where big, centralized bets fall short.
Assume The Volcano Erupts. Then Plan Backward.
Hope is not a strategy. Assume outages, throttling, and higher latency will happen. Assume costs will climb during crunch periods. Then design your practice to survive those conditions without missing a deadline.
A Practical Playbook For Law Firms
1) Map Your Dependencies
- List every workflow that uses AI: research, drafting, review, intake, discovery, analytics.
- For each, note the vendor, model family, cloud region, and failover options.
- Identify single points of failure. If one provider goes dark, what stops?
2) Build Redundancy
- Adopt multi-vendor and multi-model setups for critical tasks.
- Keep a local or on-prem fallback for key workflows (templates, precedent banks, search tools).
- Use smaller, efficient models for routine work to reduce compute dependence during peaks.
3) Set Speed Budgets
- Define maximum acceptable response times by task (e.g., 30 seconds for a clause rewrite, 5 minutes for a memo outline).
- If a tool exceeds the budget, your runbook should shift to a backup model or manual process.
- Track latency daily. Spikes are early warning signs.
4) Pressure-Test Your Vendors
- Ask for documented uptime during peak global events and major releases.
- Confirm their GPU allocation, regional distribution, and automatic failover.
- Demand clear SLAs with credits that actually matter and incident response timelines.
- Clarify who pays if your fixed-fee matter goes upside down due to their outage.
5) Get Serious About Business Continuity
- Create outage playbooks for litigation, deals, and investigations.
- Run quarterly drills: simulate a 24-hour AI outage during an active matter.
- Prebuild manual workflows and checklists. Keep forms, clause libraries, and style guides offline.
- Pre-assign human review teams to step in when tools slow down.
6) Price With Reality, Not Hope
- Model alternative fee matters with stress-tested assumptions for AI latency and cost spikes.
- Track the time spent verifying AI output. If verification never shrinks, your ROI depends on process, not tech.
- Scope contracts to include AI-related delay clauses where ethical and commercially acceptable.
7) Optimize For "Good Enough" Models
- Use smaller models where quality is sufficient. Save the heavy models for high-stakes analysis.
- Cache repeat prompts and outputs. Store reusable research and arguments to cut calls to external APIs.
- Bring retrieval and vector databases closer to your team to reduce network drag.
8) Train Your People
- Teach lawyers and staff how to switch tools fast, verify output efficiently, and spot model degradation.
- Standardize prompt patterns and review protocols so teams can swap models without losing quality.
If you're building team capability, consider structured programs that focus on practical workflows and verification habits. See role-based AI training options.
Questions To Put In Every AI RFP
- What are your peak-hour latency guarantees, region by region?
- How many GPU providers and regions backstop your service?
- What happens to my data if you throttle or go dark?
- What's your incident response target for legal customers on deadline?
- Show historical uptime and performance during major outages (provider-agnostic).
Operational Drills To Run Quarterly
- A 12-hour compute throttle where model responses slow by 50%.
- A full AI outage during e-discovery production (shift to backup workflows).
- A surge-pricing event that doubles inference costs mid-matter.
- A region-specific cloud incident that knocks out one vendor.
The Calm Before The Ash
The legal sector is acting like Pompeii - living under a volcano and assuming silence means safety. The danger isn't just hallucinations. It's the hard constraints of energy, hardware, and time.
Plan as if outages and slowdowns will hit on your worst day. Diversify your stack. Keep analog backups. Price with skepticism. And don't bet your client service on infrastructure you can't control.
AI can still help you deliver faster, smarter work - if you build for fragility, not perfection. Start now. Before the ash falls.
Your membership also unlocks:
