Construction AI BriefSubscribe →

Guide · updated 2026.11.21

Will AI compute shortages make your project software unreliable?

Yes. AI features in construction software are less reliable than the platforms they sit inside. Claude went down twice on November 13, 2026, a Microsoft 365 outage broke Claude's and ChatGPT's Outlook connectors at the same time, and the power backlog behind the compute shortage now runs past 2027. Here's what that means for your submittal, RFI, and estimating tools.

Yes. Plan on the AI features in your project software being less reliable than the software around them, for at least the next two years. Claude went down twice on November 13, 2026. A Microsoft 365 outage the same week broke the Outlook connectors in both Claude and ChatGPT. And the power-equipment backlog behind the compute shortage now runs into 2027 and beyond. Every submittal-review, RFI-drafting, or takeoff feature your construction software vendor built on those services inherits that fragility, one layer removed.

Updated November 21, 2026: added the November outage cluster, Anthropic's IPO-filing disclosures on capacity and power, and current transformer lead times. Also corrected our earlier framing: not every AI outage is a capacity outage, and the difference changes what you should ask your vendor.

Is the compute shortage actually real, or is this a vendor excuse?

It's real, and the strongest evidence now comes from a securities filing, not a status page. Anthropic's draft IPO filing, as reported in October 2026, shows about $450 billion in "massive, noncancelable" cloud commitments over the next decade, mostly with Amazon and Google. It also says the company sought more capacity than the hyperscalers would supply (24/7 Wall St.). That sits on top of the May deal to rent all of SpaceX's Colossus 1 data center, more than 220,000 Nvidia GPUs and 300 megawatts (CNBC; Anthropic).

All of that spending still hasn't stopped the outages. Claude logged its 164th documented disruption of the year on August 5 (Tech Times). On September 3, ChatGPT, Claude, and Grok all went down in the same window (Axios; The Register; BleepingComputer). Users that fall saw the message "Due to unexpected capacity constraints, Claude is unable to respond to your message" (StatusGator).

What's actually been breaking since September?

The failures have kept coming, but they've changed shape. The outages are less about raw model capacity now and more about everything wrapped around it: sign-in, connectors, agents.

Date (2026)ServiceWhat brokeStated cause
Sept 29ClaudeErrors across Claude, Claude Code, Cowork, and the API for about 36 minutes; single sign-on and file uploads failed after that (9to5Google)None given
Nov 11ChatGPTNew Workspace Agent runs blocked for about 9 hours 45 minutes, flagged as major (StatusGator)Not stated
Nov 12–13Claude and ChatGPTMicrosoft 365 connector and Claude for Outlook down for about 24 hours; OpenAI's Microsoft 365 connector also hit (AppleScoop; StatusGator)Microsoft Exchange Online gateway bug, not AI capacity
Nov 13ClaudeSign-in failure 09:58–11:58 UTC, then a broader claude.ai, API, and Claude Code outage 13:09–15:38 UTC; user reports peaked at 6,404 an hour (Instatus; DEV)None published
Nov 18ClaudeGoogle Docs connector editing failed; reading still worked (UptimeRobot)None published

Two details from November 13 matter for anyone running AI workflows. During the first outage the Claude API stayed up, so teams with a backup API key could keep working. The second outage took the API down too, and scheduled tasks ran more than two hours late or were skipped (DEV). If your vendor runs overnight batch jobs on one of these providers, like nightly submittal-log updates or morning RFI summaries, a skipped run can go unnoticed until someone asks where the report went.

Are all of these outages caused by the compute shortage?

No. Our earlier version of this page blurred that, so here's the straight version. September's "capacity constraints" errors were explicitly about compute. Most November incidents came with no published cause. The November 12–13 connector failure traced to a bug in Microsoft's Exchange Online gateway servers, made worse by extra load and capacity that was already offline (AppleScoop).

For a contractor, the cause matters less than the pattern. An AI feature in your software depends on a longer chain than the software itself: the model provider, its compute supply, its sign-in system, and the Outlook or SharePoint connector it uses to read your email. Any link in that chain can break your RFI assistant while Procore or your document system stays up.

Does this affect construction-specific software, or just chatbots?

It reaches construction tools too. Submittal copilots, RFI drafting assistants, scheduling agents, and estimating tools mostly don't run their own models. They call capacity from a frontier lab's API or from Google Cloud, AWS, or Azure. A construction software vendor is one account among thousands, with no structural reason to get priority over a lab's largest direct customers.

That shows up three ways. The feature slows down at peak hours. Requests get rate-limited or queued. Or the vendor quietly routes traffic to a cheaper, less capable model without telling you accuracy dropped. The first two are documented at the provider level. The third is hard to prove from the outside, which is why you should ask about it directly. For a worked example of a vendor dependency going wrong, see our coverage of OpenAI shutting off its Sora API the day before a Claude outage.

When will the compute shortage ease?

Not on a software timeline. The binding constraint is now power equipment and grid connections, not chips. Generator step-up transformer lead times passed 160 weeks in the first quarter of 2026, up from a 143-week average in 2024, according to Wood Mackenzie figures reported by Reuters (Wealthion). PwC analysts put lead times for high-capacity units at up to four years (pv magazine USA). Anthropic's own filing warns that power delays could slow the AI buildout, and that it "may bear these costs even if the underlying capacity is delayed or never realized" (Semafor).

How much of the pipeline is actually slipping is disputed. Sightline Climate estimated in February that 30–50% of the roughly 16 GW of U.S. data center projects planned for 2026 would be delayed to 2027 or later, or never built (E&E News). SemiAnalysis argues that figure overstates the problem because the slippage is concentrated in early-stage announced projects, while its hyperscaler build forecasts moved only a few percent (SemiAnalysis). Both camps agree that real projects are slipping and that the bottleneck is power. Neither expects it to clear in 2027. For the construction side of that backlog, see our guide to the data center construction boom and our coverage of Anthropic's filing on power delays and grid-upgrade costs.

What should you actually ask your AI vendor?

Four questions. How a vendor answers them tells you more than any feature demo:

  1. What happens to this feature under load? A rate limit, a queue, or a silent switch to a smaller model? Ask which one, and get it in writing if your workflow depends on the answer.
  2. Does your SLA cover the AI feature, or just the core platform? Most SLAs guarantee uptime for the base application. Few extend it to the AI layer, and the gap is usually in the fine print.
  3. Do you have a second model provider you can fail over to? November 13 showed that a backup API path kept some teams working while the consumer app was down. A vendor that's tied to one provider with no fallback goes down whenever that provider does.
  4. Which third-party connectors does the feature depend on? If your RFI assistant reads from Outlook or SharePoint, a Microsoft outage takes it down even when the AI provider is healthy, which is exactly what happened on November 12.

What this means for your workflow today

None of this is a reason to drop AI features that already save real hours on submittal prep, RFI drafting, or takeoff. It's a reason to stop treating them like email or your document repository. Keep anything time-critical, like a bid-day number or a same-day RFI response with contractual teeth, on a workflow a person can finish by hand. Check that overnight AI jobs actually ran before you rely on their output. Use the AI feature for the volume of work where a bad afternoon costs you some convenience, not a deadline. For how the same scarcity shows up on your invoice, see our guide to the real cost of AI in construction software. For what happens contractually when an AI tool fails at the wrong moment, see who is liable when construction AI makes a mistake.

Construction AI Brief covers what's actually shipping, and what's actually breaking, for GCs, trade subs, and estimators, three times a week. Subscribe at constructionaibrief.com.

Frequently asked questions

Is the AI compute shortage real, or is it a vendor excuse for a worse product?

It's real, and Anthropic's own IPO paperwork says so. Its draft filing, as reported in October 2026, shows about $450 billion in noncancelable cloud commitments over the next decade and says the company sought more capacity than the hyperscalers would supply. A company doesn't lock in commitments that large for a problem it doesn't have.

Are recent AI outages caused by the compute shortage?

Some are, but not all, and that matters. In September users saw explicit 'capacity constraints' errors. Most November 2026 incidents had no published cause, and one, the November 12–13 Microsoft 365 outage that broke Claude's Outlook connector, came from a Microsoft gateway bug. Either way, the AI feature in your software fails more often than the platform around it.

Does this affect construction-specific software, or just chatbots like Claude and ChatGPT?

It reaches construction tools too, one layer removed. Submittal-review copilots, RFI drafting assistants, and estimating agents mostly call the same few model providers on the back end. A construction software vendor is one account among thousands competing for that capacity and inherits every upstream outage.

When will the compute shortage ease?

Not soon. The constraint is now power infrastructure, not chips. Generator step-up transformer lead times passed 160 weeks in early 2026, per Wood Mackenzie data reported by Reuters. Anthropic's own filing warns that power delays could slow the buildout. That puts relief on a utility's timeline, which is measured in years.

What should I actually ask my software vendor about this?

Ask four things: what happens to the AI feature under load (rate limit, queue, or silent switch to a cheaper model), whether the SLA covers the AI layer or only core uptime, whether the vendor has a backup model provider, and which third-party connectors (Outlook, SharePoint) the feature depends on. A vendor that can't answer the backup-provider question hasn't planned for an outage.

Should this stop me from adopting AI features in my project software?

No. The time savings on submittal prep, RFI drafting, and takeoff are real. It's a reason not to route anything time-critical, like a bid-day deadline or an RFI response with contractual consequences, through an AI feature that has no manual fallback.

Project
Construction AI Brief
Dated
2026.11.22
Sheet
1 / 1
Rev
A
Published independently · constructionaibrief.com · © 2026Facebook·Privacy·About