Automating a $50 Billion Industry With a Single Rack

Property management is a $50 billion per year industry in the United States alone. If you look at Carpathia Capital’s chart of industries most likely to be automated by AI, property management sits right there in orange—not yet in the deep red of "already disrupted," but clearly in the zone of "this is happening now."

TIDY has already shipped the product to automate it. We’re not waiting for AGI or some future breakthrough. The AI models that exist today are sufficient. What we wanted to figure out—and what this post is about—is: what does the compute actually cost to replace $50 billion of human labor? And how does it all fit in a single server rack by 2028?

We’re going to walk through this using short-term rentals as our worked example— it’s the segment we know best—and then extrapolate to the full industry. First, the setup cost: what it takes to get a property fully live and AI-managed. Second, the cost to serve: the ongoing weekly token spend for guest operations, pro coordination, maintenance, compliance, and everything else. Third, the industry scale: what happens when you multiply by every property in the country, and the world. And finally, the hardware: how many GPUs, how many racks, how many dollars.

Our baseline assumption: a typical short-term rental with one turnover per week. Some properties turn twice or three times a week, some every other week—but one per week gives us a clean ballpark.

Why Tokens Are the Right Way to Think About This

A token is roughly ¾ of a word. Every time an AI reads context, reasons about a problem, or generates a response, it consumes tokens. Today, frontier model pricing sits around $3–15 per million input tokens and $10–75 per million output tokens, depending on the model. But these costs have been falling 10x roughly every 18 months—and that trend shows no sign of stopping.

So when we say "this task costs 50,000 tokens," what matters isn’t today’s price. What matters is the scale of the work. A 50K-token task is fundamentally small. A 5M-token task is fundamentally large. The dollar cost will keep shrinking, but the token footprint tells you what’s easy versus what’s hard.

Part 1: The Setup Cost

Getting a short-term rental property fully live and AI-managed

When a property owner comes to TIDY to get their short-term rental fully managed, there’s a one-time setup process. Let’s break down each step and estimate the token cost.

1. Analyze the Listing

The AI needs to pull in the listing from Airbnb, Vrbo, or wherever it lives, understand the property layout, amenities, house rules, and compare everything against best practices. What’s the listing missing? What could be better? Are the descriptions accurate and optimized?

The listing text itself is maybe 2,000–5,000 tokens. Best practices context and comparison frameworks add another 10,000–20,000 tokens. With a couple of reasoning loops where the AI thinks through recommendations, you’re looking at:

Listing analysis: ~50,000–100,000 tokens

That includes input tokens for reading the listing and best practices, plus output tokens for generating a detailed analysis and recommendations. At today’s prices, that’s roughly $0.15–$1.00. Trivial.

2. Photo Review & Cleanup

This is one of the more interesting setup tasks. The AI reviews listing photos to identify issues—bad lighting, clutter, poor angles—and can either suggest improvements or use generative AI to enhance them.

Photo analysis uses vision models. Each image is roughly 1,000–2,000 tokens to process (depending on resolution). A typical listing has 20–40 photos. For each photo the AI needs to assess quality, suggest improvements, and potentially generate enhanced versions.

Photo review: ~80,000–160,000 tokens (vision)
Photo enhancement: ~$0.50–$2.00 in image generation API calls

The image generation side (calling something like Google’s Imagen or similar) is priced per image rather than per token, but it’s still on the order of a few cents per image. For a full listing refresh, you might spend $0.50–$2.00 on photo enhancement total.

3. Analyze Local Regulations

Short-term rental regulations vary wildly by city, county, and state. The AI needs to research what’s required: permits, tax registrations, occupancy limits, noise ordinances, HOA rules, and more.

This involves pulling in regulatory documents (which can be lengthy—some municipal codes run thousands of pages) and extracting the relevant sections. The AI doesn’t need to read the entire code—it retrieves the relevant sections, typically 20,000–50,000 tokens of regulatory text, then reasons about applicability.

Regulatory analysis: ~100,000–200,000 tokens

This is one of the heavier setup tasks in terms of tokens because of the volume of source material. But it’s still a one-time cost, and the regulatory knowledge can be cached and reused across properties in the same jurisdiction.

4. Generate & File Regulatory Paperwork

Once regulations are understood, the AI can help generate permit applications, tax registrations, and compliance documentation. This is mostly form-filling and document generation—relatively light token work.

Regulatory filings: ~20,000–50,000 tokens

5. Set Up Dynamic Pricing

Configuring a pricing engine—connecting to market data, setting base rates, seasonal adjustments, event-based pricing, minimum stays. This is mostly software, not AI reasoning. The AI’s role is light: understanding the owner’s pricing preferences and translating them into configuration.

Pricing setup: ~10,000–20,000 tokens

6. Connect Financial Systems

Linking bank accounts, payment processors, and accounting systems. This is pure software integration—API calls, OAuth flows, webhook setup. The AI might have a brief conversation to confirm which accounts to connect, but the actual work is code, not reasoning.

Financial setup: ~5,000 tokens (essentially zero)

7. Set Up Smart Locks & Access

Connecting smart locks, generating access codes, setting up automated code rotation for guests. Again, this is primarily software—API integrations with lock providers. The AI helps configure the rules (when to send codes, how to rotate them), but the heavy lifting is traditional software.

Smart lock setup: ~5,000–10,000 tokens

8. Configure the TIDY Platform

This is setting up the scaffolding in TIDY itself—creating the property profile, cleaning checklists, maintenance schedules, pro assignments, notification preferences, and agent configurations. It’s a combination of software configuration and AI-assisted customization.

Platform configuration: ~20,000–40,000 tokens

9. Analyze Historical Messages & Extract Patterns

This is potentially the most token-heavy setup task. If the property has existing operations, the AI needs to ingest historical guest messages, pro communications, and owner correspondence to understand patterns—what questions guests typically ask, how the owner likes things handled, what issues come up repeatedly.

A property with a year of operations might have thousands of messages. Even summarizing and extracting key concepts from this history could require processing significant volumes of text. However, this can be done with summarization chains—process messages in batches, extract key themes, then synthesize.

Historical analysis: ~200,000–500,000 tokens

The range here is wide because it depends entirely on how much history exists. A brand-new listing has zero history to analyze. A property with years of guest reviews and pro communications could be at the high end. But even at 500K tokens, we’re talking about $1.50–$7.50 at today’s prices.

10. The Sales Conversation

Finally, there’s the back-and-forth conversation with the property owner during setup. Understanding their goals, preferences, deal-breakers, and specific requirements. This is a multi-turn conversation that might span several sessions.

Sales & onboarding conversation: ~50,000–150,000 tokens

Adding It All Up: Total Setup Cost

Plus roughly $0.50–$2.00 in image generation API calls for photo enhancement.

At today’s frontier model pricing (let’s use $10/M input, $30/M output as a blended average for a capable model), and assuming roughly a 3:1 input-to-output ratio:

Total one-time setup cost: roughly $5–$20 per property

Five to twenty dollars in AI compute to do what traditionally requires hours of human work—analyzing listings, researching regulations, reviewing photos, configuring systems, and having detailed onboarding conversations.

Part 2: The Ongoing Cost to Serve

What it costs in tokens to operate a property week after week

Setup is a one-time cost. The real question is: what does it cost to actually run the property, week in and week out? This is where a property manager spends their time—and where the token economics tell the most interesting story.

We’ll assume one turnover per week (~52 guests per year) and break everything into per-turn costs and amortized ongoing costs.

Guest Operations: Inquiries, Bookings & Communication

As new inquiries come in, the AI needs to evaluate them against your house rules, availability, guest history, and pricing. Much of this is handled by software—calendar lookups, rule engines, pricing algorithms—so the token budget is really just the AI applying judgment to edge cases and generating personalized responses.

Guest inquiry & booking: ~5,000–15,000 tokens per guest

During the stay itself, the token spend is modest. The average guest sends maybe 2–3 messages—check-in questions, local recommendations, a minor issue. These are mostly solved with text inputs and outputs. Each exchange runs about 3,000–7,000 tokens including context.

Guest communication during stay: ~10,000–20,000 tokens per guest

Roughly 1 in 50 guests has something that escalates—"show me the video," "I need to see what’s going on"—requiring video analysis or a more complex multi-step resolution. When that happens, it might be 100,000–200,000 tokens. But amortized across all guests:

Escalation handling (amortized): ~2,000–4,000 tokens per guest

Pro Coordination: The Biggest Ongoing Token Spend

This is honestly the most token-heavy part of ongoing operations, and it’s worth explaining why.

For every turnover, there’s messaging with your cleaning pro: scheduling confirmation, reminders, hearing about any issues they found, answering questions. Pros often ask things they could find by looking at their checklist—but they frequently don’t, and we don’t always get feedback to know what’s unclear. So there’s a steady stream of back-and-forth messaging. On average, maybe 3–5 message exchanges per turn, each consuming tokens for context and response.

Pro messaging per turnover: ~15,000–30,000 tokens

Where the token budget really spikes is when you need a new pro. Your regular cleaner can’t make it, or you’re replacing a provider, and now the AI needs to search through potentially hundreds of candidates, evaluate profiles, check availability, negotiate rates, and onboard someone new.

A lot of the initial filtering can be done by structured database queries and hard-coded rules—location, availability, ratings, price range. But the AI needs to analyze the filtered shortlist, assess fit, and handle the outreach conversation. When this happens, it’s a 200,000–500,000 token event. But it only happens maybe 4–8 times per year for a typical property.

New pro search (amortized): ~5,000–15,000 tokens per week

Post-Clean Review: Photos & Quality Verification

After each cleaning, you’re reviewing before-and-after photos to verify quality. A typical turn generates 5–20 photos that need analyzing—checking that everything matches the checklist, spotting issues, confirming the property is guest-ready.

Occasionally there’s a 5-minute video walkthrough to review. Photo analysis runs about 1,500–2,000 tokens per image through a vision model. Video is heavier when it happens but doesn’t happen every turn.

Post-clean photo/video review: ~15,000–40,000 tokens per turn

Maintenance & Restocking

Maintenance is intermittent—a plumber here, an HVAC call there, a handyman for a broken shelf. Each maintenance job has roughly the same token profile as a cleaning coordination: dispatching, messaging the pro, reviewing results. Restocking (toiletries, linens, cleaning supplies) follows a similar pattern—detect what’s needed from checklists or photos, trigger orders or pro tasks.

For a typical short-term rental, you might have maintenance pop up every 2–4 weeks and restocking every 1–2 weeks. These are in the same order of magnitude as other job types.

Maintenance (amortized): ~5,000–15,000 tokens per week
Restocking: ~5,000–10,000 tokens per week

Compliance & Regulatory Upkeep

The initial regulatory analysis happens at setup, but there’s ongoing compliance work: periodic filings, tax payments, permit renewals, occupancy reports. These compliance systems are often old school—log into a government portal, fill out a form, send a check. If the AI is using web browsing to handle these, that consumes tokens for page content, navigation decisions, and form filling.

A typical compliance filing is roughly the equivalent of 3 minutes of web browsing—maybe 20,000–50,000 tokens per session. These happen monthly or quarterly.

Compliance upkeep (amortized): ~5,000–10,000 tokens per week

There are also noise alerts—a noise sensor triggers, and the AI needs to message the guest and handle the response. This is mostly software-driven (sensor triggers the workflow) with a small token budget for crafting the message and managing any back-and-forth.

Noise alerts & compliance triggers (amortized): ~2,000–5,000 tokens per week

Financial Monitoring & Owner Communication

Monitoring financial transactions, categorizing income and expenses, spotting anomalies, generating owner reports. Much of this is a software problem—database queries, accounting rules—with the AI handling edge cases and generating human-readable summaries.

Financial monitoring: ~3,000–8,000 tokens per week
Owner updates & communication: ~5,000–10,000 tokens per week

Marketing, Pricing & Listing Optimization

Dynamic pricing adjustments, listing syndication across platforms, upsell opportunities— this is overwhelmingly a software problem. Pricing algorithms run on structured data. Listing syndication is API calls. The AI’s role is occasional: maybe tweaking listing copy seasonally, suggesting pricing strategy changes, or identifying upsell opportunities.

Marketing & pricing: ~2,000–5,000 tokens per week

Last-Minute Changes & Exceptions

Last-minute cancellations, guest reschedule requests, shuffling pro assignments—these create incremental work but don’t happen every turnover. Maybe 1 in 4 weeks you have some kind of schedule disruption that requires the AI to recoordinate. When it happens, it’s a 30,000–60,000 token event. Amortized:

Schedule disruptions (amortized): ~3,000–8,000 tokens per week

Adding It All Up: Weekly Cost to Serve

Scaling that up:

• Per month: ~330K–780K tokens
• Per year: ~4.3M–10.1M tokens

The Annual Token Budget for One Property

Now we can put the full picture together—setup plus a year of operations:

These numbers come from production. We use blended frontier model pricing at roughly $15 per million tokens (a 3:1 input-to-output ratio at $10/M input and $30/M output), plus image processing costs.

The wide range reflects real-world variance. We’ve measured per-property token consumption as low as 5 million tokens per year for well-established properties with routine operations and optimized prompts. But we err on the side of the high end of the range—up to 10x—to increase accuracy and account for complex edge cases, heavy reasoning tasks, regulatory research, difficult pro searches, and properties with unusual requirements.

We expect these costs to decrease over time. Better models, tighter scaffolding, smarter automation, and growing confidence in lighter-weight approaches for routine tasks all push token consumption lower. The low end of our range today is where we expect the average to settle.

Compare That to a Traditional Property Manager

A short-term rental generating $50,000/year in revenue with a traditional property manager at 10–15% fees pays $5,000–$7,500 per year in management costs.

The AI token cost to do 99% of the same work? In production, we’ve measured as low as $75 in Year 1. Our conservative budget—erring high for accuracy—is up to $2,150. Even at the conservative end, the AI compute cost is less than half of traditional property management fees.

At our optimized production levels, that’s a 25–100x cost advantage in raw compute. Even at our most conservative budget, it’s a 2–3x advantage—and improving fast. At TIDY’s 3.9% fee ($1,950/year on a $50K property), there’s massive margin to deliver better service at a fraction of the industry’s traditional cost.

And remember: token costs are falling roughly 10x every 18 months. By 2028, the same operational workload might cost $7–$20 per property per year in AI compute. Multiply that across a $50 billion industry and the economics are staggering.

Where the Tokens Actually Go

Looking at the ongoing cost to serve, a clear pattern emerges. The biggest token consumers are:

1. Pro coordination (~40% of weekly tokens)
   Regular messaging, new pro search, quality verification. This is the largest single category because it involves the most human-like reasoning—understanding context, negotiating, evaluating quality, handling exceptions. Pros often don’t check their checklists and instead ask questions, which means more back-and-forth than you’d expect from a pure software standpoint.
2. Guest communication (~25% of weekly tokens)
   Inquiries, during-stay messages, and the occasional escalation. Most of this is straightforward text exchanges, with rare spikes for complex issues.
3. Everything else (~35% combined)
   Maintenance, restocking, compliance, financial monitoring, owner updates, marketing, and schedule disruptions. Individually small, collectively significant, and heavily assisted by traditional software. The AI fills in the gaps that software can’t handle alone.

What Most People Get Wrong

When people hear "AI property management," they imagine a massive compute requirement—some kind of always-on superintelligence monitoring every aspect of every property. The reality is far more prosaic.

Most of what a property manager does is coordination work—scheduling, messaging, checking, confirming—and the average turn generates maybe 100K–200K tokens of AI work. That’s roughly 75,000–150,000 words of reading and writing. For context, that’s about the length of one or two novels. Once a week. For an entire property.

A huge portion of the work isn’t even AI at all—it’s traditional software. Calendar systems, pricing algorithms, payment processing, smart lock APIs, database queries for pro matching. The AI is the reasoning layer on top, handling the judgment calls, the natural language communication, and the visual inspection. The scaffolding underneath is just code.

Part 3: The Industry-Scale Token Budget

What it costs to AI-manage every short-term rental in the US—and the world

We’ve established that a single property costs roughly 5–11 million tokens per year to manage with AI. Now let’s zoom out. What does the total token demand look like if you wanted to AI-manage every short-term rental?

How Many Properties Are We Talking About?

The US has approximately 1.8 million unique short-term rental properties (per AirDNA and industry data). Some sources count up to 2.5 million listings across platforms, but that includes cross-listing on Airbnb, VRBO, and Booking.com simultaneously. We’ll use 1.8 million unique properties.

Globally, Airbnb alone has about 7.7 million active listings. Add VRBO (2M+), Booking.com’s alternative accommodations, regional platforms, and direct-booking properties, and adjust for cross-listing overlap, and the global figure lands around 10 million unique short-term rental properties.

US Token Demand

Using our production range—from optimized deployments at the low end to conservative 10x budgets at the high end:

Global Token Demand

Compare That to What the Industry Spends Today

Let’s put these numbers in context. The US property management industry as a whole— short-term rentals, long-term rentals, commercial, HOA management—is roughly a $50 billion per year industry. The short-term rental slice we’ve been analyzing represents about $2–$3 billion of that (10–15% management fees on ~$20 billion in US STR revenue). Globally, STR management fees run $12–$23 billion per year (on a $125–$155 billion market).

The AI token cost to replace the STR management layer, based on our production data:

• US STR (optimized): $135M–$300M (vs. $2–$3B traditional) — a 7–22x cost advantage
• US STR (conservative 10x): $1.4B–$3B — roughly matching traditional costs today, but falling fast
• Global STR (optimized): $750M–$1.7B (vs. $12–$23B traditional) — a 7–30x cost advantage

Even at our most conservative production budget, the AI cost is at parity with traditional management—and token costs are falling roughly 10x every 18 months. By 2028, even the conservative end drops to a 7–20x advantage. At optimized levels, it’s 70–200x.

And that’s just the STR segment. The same token economics apply to the broader $50B US property management industry. The tasks are similar—coordination, communication, scheduling, compliance, quality control—they just have different nouns. The token math scales linearly.

Optimized vs. Conservative: What’s the Difference?

Our "optimized" numbers are what we’ve achieved in production on well-established properties with routine operations, tight prompts, and efficient scaffolding. Most property management tasks are straightforward—scheduling, simple messaging, photo review, financial monitoring—and run efficiently on capable models.

Our "conservative" budget is up to 10x higher. We err on this side to increase accuracy across the full range of real-world conditions: complex regulatory analysis requiring multi-step legal reasoning, difficult pro searches with nuanced candidate evaluation, guest escalations requiring judgment, properties with unusual requirements, and the long tail of edge cases that don’t show up in averages.

In practice, most properties run much closer to the optimized end. We expect the conservative budget to tighten over time as models improve, our scaffolding gets more efficient, and we gain confidence in lighter-weight approaches for tasks that currently get the heavy reasoning treatment.

Putting It in Perspective

To AI-manage every short-term rental in the United States for a full year—setup, operations, guest communication, pro coordination, maintenance, compliance, everything— requires somewhere around 9 to 200 trillion tokens, depending on where you sit in our production range.

That sounds like an enormous number. But consider: the major AI labs are already processing hundreds of trillions of tokens per month across all their customers. Even at our conservative 10x budget, the entire US short-term rental industry’s management needs represent a modest slice of current global AI capacity.

The compute exists. The models are capable. The economics work. What’s left is execution—the software scaffolding, the integrations, the refinement of every edge case. That’s what TIDY has been building for 12+ years.

Part 4: GPUs, Hardware, and Real Dollar Costs

How many GPUs does it take to manage every rental on Earth—and what does it actually cost?

We’ve established the token budget. Now let’s translate that into physical hardware. How many GPUs would you actually need? What would they cost? And what if you just bought the tokens from Anthropic instead?

First: How Fast Can Today’s GPUs Generate Tokens?

For a large, capable model (think 670B+ parameters—something in the class of DeepSeek-R1 or comparable to the frontier models used for complex reasoning), here’s what current and upcoming NVIDIA hardware can do:

• NVIDIA B200 (Blackwell): ~3,750 tokens per second per GPU on a 670B model. A DGX B200 (8 GPUs) hits ~30,000 TPS total. Cost: ~$45,000 per GPU.
• NVIDIA GB200 NVL72 (rack-scale Blackwell): With NVLink connecting 72 GPUs, effective per-GPU throughput improves to ~4,500 TPS. The full rack does ~1.5M TPS on a 120B model, less on larger models. Cost: ~$3 million per rack.
• NVIDIA GB300 NVL72 (Blackwell Ultra): 45% faster than GB200 on DeepSeek-R1—roughly ~5,800 TPS per GPU. Available now.
• NVIDIA Vera Rubin (shipping H2 2026): NVIDIA claims a 10x reduction in inference token cost and 5x raw FLOPS over Blackwell. Conservative estimate: ~15,000–25,000 TPS per GPU on a 670B-class model. 288GB HBM4 per GPU. Draws ~1,800W. Price: likely $50,000–$70,000 per GPU.

Translating Token Demand Into GPU Count

Our global STR token demand spans a wide production range: 80 trillion tokens per year at our optimized levels, up to 800 trillion at the conservative end. Let’s use the optimized midpoint for the hardware analysis—the conservative end simply scales linearly (10x the GPUs, 10x the racks):

• Average throughput (optimized): 80T tokens ÷ 31.5M seconds/year ≈ 2.5 million tokens per second
• Peak throughput (2x for daytime clustering): ~5 million TPS
• With 50% headroom over peak: ~7.5 million TPS capacity needed

Property management isn’t uniform around the clock—turnovers cluster in mornings, guest messages spike in evenings, compliance work happens during business hours. But it does span global time zones, which smooths things out. A 2x peak-to-average ratio with 50% headroom is conservative.

Option 1: Build It With Blackwell (Available Now)

Using B200 GPUs at ~3,750 TPS per GPU on a large capable model:

Twenty-eight racks at our optimized production levels. At our conservative 10x budget, that’s 280 racks—still a small footprint in a single data center. For context, a single hyperscaler data center might house thousands of racks.

Option 2: Wait for Vera Rubin (H2 2026)

Using projected Vera Rubin throughput at ~20,000 TPS per GPU:

Five racks at optimized levels, 50 at our conservative budget. Either way, you could fit the entire world’s short-term rental AI management infrastructure in a single server room. The annual cost—hardware, power, operations, everything—is about what a mid-size property management company spends on salaries.

Option 3: Just Buy the Tokens — Claude Opus 4.6 Pricing

Most companies won’t build their own GPU clusters. They’ll buy tokens from an API provider. Let’s price this out using Claude Opus 4.6—Anthropic’s most capable model—at full retail pricing. This is deliberately the most expensive option to show the ceiling.

Opus 4.6 pricing: $5 per million input tokens, $25 per million output tokens. Assuming a 3:1 input-to-output ratio (typical for property management—lots of reading context, shorter responses):

Prompt caching is particularly effective for property management because many operations share context—the same property details, checklist templates, house rules, and regulatory frameworks get loaded repeatedly. Anthropic’s caching drops input costs by 90% for cached tokens ($0.50/M instead of $5/M), and most property context qualifies.

But You Don’t Need Opus for Everything

Here’s the thing: most property management tasks don’t require the most powerful model on the planet. Scheduling confirmations, routine guest messages, photo reviews, and financial categorization work great on smaller, cheaper models. You really only need Opus-class reasoning for complex tasks: regulatory analysis, difficult guest escalations, nuanced pro evaluation, and tricky judgment calls.

A realistic model mix might be:

• ~70% of tokens on a fast, efficient model (like Haiku-class at $0.25/$1.25 per MTok) — routine scheduling, simple messages, standard reviews
• ~25% on a mid-tier model (like Sonnet-class at $3/$15 per MTok) — pro coordination, guest communication, photo analysis
• ~5% on Opus-class ($5/$25 per MTok) — regulatory analysis, complex escalations, new pro search evaluation

The Full Comparison

Let That Sink In

The US property management industry spends $50 billion per year on human labor. Just the short-term rental slice—$2–$3 billion—could be replaced with AI. At our optimized production levels with smart model routing and batch processing, the US cost is $18–$23 million. Even at our conservative 10x budget, it’s $180–$230 million—still a 10x cost advantage.

At optimized levels, a Vera Rubin cluster could AI-manage every short-term rental in America for $2–$3 million per year with five server racks. At the conservative end, that’s 50 racks and $20–$30 million—still a fraction of the industry’s cost.

Globally? $11–$140 million per year depending on where you are in the production range. Even the high end is a rounding error compared to $12–$23 billion in traditional management fees. And that’s just the STR segment—the same economics apply across the full $50 billion US property management industry.

Important Caveats

These GPU calculations assume you’re running open-source or self-hosted models. You can’t run Claude Opus on your own hardware—it’s a proprietary model available only through Anthropic’s API. The GPU cluster numbers represent what you’d need if running an equivalently capable open model (like DeepSeek-R1 or future open alternatives).

The API pricing path is the realistic one for most companies. The GPU path illustrates the underlying physics: this workload is not computationally large. The gap between the ~$12M self-hosted cost and the ~$95–330M API cost represents the model provider’s margin, their R&D costs, and the premium for the best models.

Also: none of this includes the software platform itself—the scheduling engine, the integration layer, the pro network, the compliance tracking, the digital twin. The tokens are just the AI reasoning layer. The scaffolding underneath is what makes those tokens useful.

A Rounding Error on Global GPU Production

Here’s perhaps the most striking way to frame all of this. NVIDIA’s Vera Rubin platform entered full production in early 2026—ahead of schedule—with volume shipments starting in H2 2026. Analyst estimates project roughly 30,000 Vera Rubin NVL72 racks shipping in 2026 (that’s ~2.16 million GPUs), ramping to ~100,000 racks (~7.2 million GPUs) in 2027.

At our optimized production levels, we need 5 racks to AI-manage every short-term rental on Earth. At our conservative 10x budget, 50 racks.

5–50 racks out of 30,000 produced = 0.017–0.17% of Year 1 Vera Rubin production

The entire global short-term rental property management workload is a rounding error on NVIDIA’s output. Even if you expand to all property types—long-term rentals, commercial properties, HOAs, the full scope of property management worldwide—and use our conservative budget, you might need 500 racks. That’s still 1.7% of Year 1 production.

The hardware constraint doesn’t exist. The compute to automate a $50 billion industry is being manufactured at a scale that makes the workload invisible. The only question is who builds the software scaffolding to put those tokens to work—and that’s been TIDY’s focus for over a decade.

Why We Expect These Costs to Decrease

Our production range is wide—up to 10x between optimized and conservative—because we deliberately err on the side of higher token budgets to increase accuracy. But we’re actively working to narrow that range downward, and multiple forces are converging to make it happen:

• Better models: Each generation of frontier models delivers better reasoning with fewer tokens. Tasks that required long chain-of-thought reasoning a year ago now complete in a single pass.
• Tighter scaffolding: Every month, we optimize our prompts, reduce unnecessary context loading, and improve the software layer that handles what doesn’t need AI at all. The scaffolding gets more efficient with every deployment.
• Smarter automation: As we accumulate production data, we identify which tasks can be handled by smaller, cheaper models—or by pure software with no AI reasoning at all. The model mix shifts toward efficiency.
• Growing confidence: Early in a deployment, we over-invest in reasoning to ensure quality. As we validate performance on specific task types, we can safely reduce token budgets without sacrificing accuracy.

The direction is clear: our conservative budget tightens toward our optimized levels, and our optimized levels continue to improve. Meanwhile, the underlying cost per token keeps falling roughly 10x every 18 months. Both curves bend the same way.

Why 99%, Not 100%—and the Human Cost of the Last 1%

We’ve been careful to say 99% automation, not 100%. This is deliberate. The last 1% isn’t a technical limitation—it’s a design choice.

If the house catches fire, do you want the AI to handle it without notifying the owner? If there’s a major insurance claim, a lawsuit from a guest, or a decision to terminate a long-standing pro relationship—these are moments where owners may genuinely want to be involved. Going much beyond 99% automation in this industry isn’t just unnecessary, it’s arguably not even desired.

So what does the remaining 1% cost in human labor?

The US property management industry employs roughly 250,000 people in the short-term rental space. At 99% automation, that’s 1%—about 2,500 peoplehandling the edge cases that need a human touch.

The Full Organization: Not Just Automation

Automation doesn’t mean there are zero humans. Even a fully AI-managed property management platform needs people—just far fewer, and doing different work.

You still need account managers and sales representatives to help property owners get set up, ease the transition from traditional management, and provide the human relationship that some owners want. You need people building and improving the software platform itself. The whole organization might run on roughly $20M in G&P labor costs.

There’s also a significant channel market: an estimated 10,000+ property managers and partners who, for various business reasons, need to be the ones using the software rather than the property owner directly. They may manage portfolios for absentee owners, handle commercial relationships, or serve markets where direct-to-owner AI management isn’t the right fit today. This partner channel needs its own sales force—maybe another $20M in labor.

At optimized production levels, call it $200 million per year, all-in, to AI-manage every short-term rental on Earth—compute, humans, sales, engineering, everything. At our conservative budget, $1–$1.3 billion. The global STR management industry currently spends $12–$23 billion. The broader US property management industry? $50 billion.

That’s a 10–115x cost reduction on STR alone, depending on where you sit in our production range. And the range is tightening downward as models improve and our scaffolding gets more efficient.

The Product Already Exists

Everything we’ve described isn’t theoretical. TIDY has shipped this product. It’s in market today. We charge 3.9% per completed job—compared to the industry average of 10–15%. The cost savings of a software-first approach make this price point possible, and we expect it to go down further as token costs continue to fall.

It took us 12 years to build the scaffolding. We don’t think it will take competitors 12 years to match—the AI tools are better now, the playbook is more visible. But the product exists today. It hasn’t saturated the market yet. There will be competitors. And within five years, any property management business that hasn’t adopted AI automation will be dead.

This requires zero innovation from AI model providers. No breakthroughs needed. No AGI required. The models we have today are sufficient. The scaffolding just needs to exist—and it does.

That’s just property management. Just one segment on Carpathia’s chart. One fraction of one day of GPU production. How many other $50 billion industries will fall to the same math?

A Personal Note

Building this has been simultaneously the most fun of my career and one of the most terrifying things I’ve witnessed. People debate what happens if we get AGI, what happens if AI can do everything. Those conversations are important.

But they’re missing the point. Sufficient intelligence is here today. A $50 billion industry can be automated with a single rack of GPUs and the right software. Not theoretically—we’ve done it. Most industries are not safe. Not from some hypothetical future superintelligence, but from the AI that exists right now, combined with years of domain-specific scaffolding. The disruption isn’t coming. It’s here.

Jevons’ Paradox: Why Cheaper Means Bigger

Here’s what makes this story even more interesting. Today, roughly 80% of the addressable market is not using a property manager at all. They manage their own rentals because traditional property management is too expensive. When you’re paying 10–15% of revenue, plenty of owners decide it’s not worth it and do everything themselves.

This is a classic case of Jevons’ Paradox: when you make something dramatically cheaper, demand doesn’t stay flat—it explodes. A 90% cost reduction doesn’t just save money for existing customers. It pulls in the vast majority of property owners who were previously priced out.

We expect to see two massive waves of new demand:

1. Rental property owners coming off the sidelines. The 80% who manage their own properties today will start adopting AI-powered management when the cost drops from thousands of dollars per year to hundreds or less. Why handle scheduling, guest messages, and pro coordination yourself when AI can do it for 3.9%?
2. Consumers entering the market for the first time. Homeowners who have never used a property manager—because property management was historically only for landlords and investors—will start using AI-powered management tools for their own homes. Cleaning scheduling, maintenance coordination, vendor management. These are problems every homeowner has. They’ve just never had an affordable solution.

The addressable market isn’t 1.8 million US properties or 10 million globally. When the cost of management drops 90%, the market could be 5–10x larger. TIDY is already positioned for this with our consumer offering—being one of the first platforms to bring the benefits of property management to regular homeowners at a price point that makes sense.

2028: One Rack to Automate an Entire Industry

The numbers above already feel absurd—five Vera Rubin racks to manage every short-term rental on Earth. But there’s a real possibility it gets even more extreme.

NVIDIA announced Groq LPU integration into the Vera Rubin platform, which is designed to deliver up to 35x more inference throughput for certain workloads compared to standard GPU inference. Groq’s Language Processing Units are purpose-built for sequential token generation—the exact workload that dominates property management AI (reading context, generating responses, reasoning through decisions).

Combine that 35x inference improvement with the ongoing algorithmic and architecture improvements in model efficiency—distillation, mixture-of-experts, speculative decoding, quantization advances—and the math starts to look like this:

• Today (Blackwell): 28–280 racks for global STR management
• H2 2026 (Vera Rubin): 5–50 racks
• 2028 (Vera Rubin + Groq + algorithmic gains): Potentially a single Feynman rack

At optimized levels, a single rack. At our conservative budget, maybe ten. Costing $10–$100 million. To AI-manage the entire global property management industry.

A $50 billion industry of human labor, replaced by $10–$100 million in hardware. A 500–5,000x cost reduction for the US alone—not in theory, not in a decade, but within 2–3 years at current trajectories.

There are reasons this might not play out exactly this way. Model quality could plateau. Groq integration might not deliver the full 35x for real-world mixed workloads. Regulatory or market dynamics could slow adoption. But from a pure token-cost and hardware perspective, the math says it’s very possible. The trend lines are not subtle—they’re exponential, and they’re converging.

The Bottom Line

Property management is a $50 billion industry. Here’s how the math works out:

• One property, one year: 5–113M tokens in production, $75–$2,150 at today’s blended pricing. We’ve run as low as the bottom of that range and budget at the top for accuracy.
• Every US short-term rental: 28–280 Blackwell racks today, 5–50 Vera Rubin racks in H2 2026
• Every STR on Earth: $11–140M/year self-hosted, $95M–$1.2B/year via API — compared to $12–$23B in traditional management fees
• By 2028: Vera Rubin + Groq + algorithmic gains could compress the entire global STR workload into 1–10 racks costing $10–$100M
• Share of global GPU production: 0.017–0.17% of Year 1 Vera Rubin output — the hardware to automate a $50B industry already exists in absurd surplus

The product already exists—TIDY shipped it. The market hasn’t saturated yet. Competitors will come. But the math is the math: a $50 billion industry of human labor can be replaced by a handful of server racks and 12 years of software scaffolding—and the costs are heading down, not up.

Take care of your properties like it’s 2026, not 1996.

Sources & References

• AirDNA US Short-Term Rental Outlook Report — 1.8M unique US properties, market performance data
• Anthropic Claude API Pricing — Opus 4.6, Sonnet, and Haiku token pricing
• NVIDIA Vera Rubin Platform Announcement — 10x inference cost reduction, HBM4, H2 2026 availability
• Inside the NVIDIA Vera Rubin Platform — Technical blog with detailed specs and performance claims
• NVIDIA GB300 NVL72 (Blackwell Ultra) — Specs and AI reasoning performance
• CoreWeave GB300 NVL72 DeepSeek-R1 Benchmarks — 6x+ performance gain on DeepSeek-R1 workloads
• NVIDIA Groq 3 LPX Technical Blog — Groq LPU integration, up to 35x inference throughput per megawatt
• NVIDIA’s $20B Groq Acquisition Analysis — Next Platform deep dive on the deal
• NVIDIA Feynman Architecture Roadmap — 3D-stacked GPUs, Rosa CPU, 2028 timeline
• Jevons’ Paradox — Economic principle: efficiency gains increase total resource consumption
• a16z: LLMflation — LLM Inference Cost Trends — ~10x cost decline per year for equivalent AI performance
• Epoch AI: LLM Inference Price Trends — Rigorous analysis showing median 50x/year decline in token costs
• Grand View Research: Short-Term Vacation Rental Market — Global market sizing ($134.5B in 2024)