Build Big, Build Small, or Not at All

Framing

MIT Technology Review's July/August 2026 issue is nominally about engineering — chips, subsea tunnels, dark matter, and cooling the planet on purpose. Its cover poses a blunt question, borrowed from a child's model kit: build big, build small, or not at all. That turns out to be exactly the question every hotel owner, general manager, and revenue leader is now facing with AI, dressed in a different industry's overalls.

This paper sets aside the parts of the issue with no bearing on travel and hospitality — dark matter detectors, geoengineered clouds, cosmic puzzles — and pulls out the eight signals that do. Read together, they describe an AI industry entering its engineering-reality phase: the easy gains are banked, the costs are becoming visible, agents are arriving as coordinated teams rather than solo chatbots, and the public is starting to push back. None of that changes what AI can eventually do for a guest journey. It changes how an operator should spend the next twelve months getting ready for it.

Each signal below is paired with a plain reading of its meaning at the property level and three concrete implications. The full twelve-month sequence follows in the roadmap section.

01. The Chip Bill Behind the Concierge Bot

SOURCE SIGNAL — “Light work” — ASML and the economics of AI-grade chipmaking.

ASML's newest lithography machine, the one keeping pace with AI's appetite for denser chips, costs roughly $400 million per unit, and only a handful of companies on Earth can build the tools — or the chips — that depend on it. That scarcity is a large part of why frontier AI remains expensive to run, and why the cost curve has not bent downward as fast as vendor marketing implies.

WHAT IT MEANS FOR TRAVEL & HOSPITALITY

Every PMS, CRS, chatbot, and revenue-management vendor selling an “AI-powered” feature is a downstream customer of this same chip economy. When compute gets scarcer or export controls tighten, the cost eventually surfaces in a software renewal notice, not in ASML's invoice.

• Read renewal notices for AI or compute surcharges before reading the feature list.

• Ask vendors directly whether pricing is exposed to GPU or foundry capacity, and how they hedge it.

• This is exactly the terrain Pertlink's Token Cost Per Guest (TCPG) metric was built to track — model cost per guest interaction, not per software seat.

02. The Grid Becomes a Revenue-Management Problem

SOURCE SIGNAL — “Stretch goals” — how data centers are learning to throttle themselves.

Data centers are increasingly asked to dial down their own power draw during grid stress rather than wait years for new power plants. Software such as Emerald AI's Conductor can cut AI compute by roughly a quarter for hours at a stretch without visibly degrading service. Utilities favor it because it defers costly new infrastructure; AI vendors favor it because it gets facilities connected faster.

WHAT IT MEANS FOR TRAVEL & HOSPITALITY

This is invisible to a hotel — until it isn't. A property leaning on always-on AI (live pricing, real-time chat, in-room voice) is leaning on infrastructure that may deliberately be throttled at the exact moments — heat waves, storms, peak demand — that also raise guest expectations.

• Ask cloud and AI vendors what happens to your service during a grid-flex event: silent slowdown, queuing, or fallback to cached logic?

• Build an “AI goes quiet” contingency into peak-period SOPs, the same way you already plan for a Wi-Fi or PMS outage.

• Vendor energy-flexibility commitments are also a genuine ESG data point worth citing in owner and franchisor reporting.

03. Agent Orchestration Arrives at the Front Desk

SOURCE SIGNAL — “Agent orchestration” — coordinated teams of AI, not solo chatbots.

The near-term shift isn't a smarter single chatbot. It's teams of AI agents working in concert — one drafts, one checks, one escalates — coordinated toward a single outcome, the way tools like Claude Code, OpenAI's Codex, and Google DeepMind's Co-Scientist now coordinate for developers and researchers. The article's own comparison is the assembly line, rebuilt for knowledge work.

WHAT IT MEANS FOR TRAVEL & HOSPITALITY

The practical opportunity for a property isn't “a better chatbot” for guest messaging. It's a small orchestrated crew — one agent drafts the reply, one checks it against brand voice and rate integrity, one flags anything that needs a person — running behind a single shared inbox.

• Pilot orchestration on a bounded, low-risk workflow first: pre-arrival messaging or review responses, not complaint resolution or rate exceptions.

• Decide which agent role, if any, is allowed to touch price — before the first agent is switched on, not after.

• This is the near-term shape of the Human Experience Orchestrator (HXO) role: someone has to own the crew, not just the tool.

04. The “AI Coworker” Trap

SOURCE SIGNAL — The Algorithm — “AI agents are not your ‘coworkers.’”

A Boston University study found that when identical AI-generated work was labeled as coming from an “AI employee” rather than a chatbot, people caught 18% fewer errors in it, and were 44% more likely to escalate a judgment call rather than take responsibility themselves. Nearly a third of managers surveyed said their companies already frame AI agents as employees; almost a quarter list them on the org chart.

WHAT IT MEANS FOR TRAVEL & HOSPITALITY

Giving an AI concierge a name badge, a “team member” bio on the website, or a seat on the org chart is not a harmless branding choice. It measurably changes how staff supervise it — more trust, less scrutiny, more buck-passing the moment something goes wrong.

• Keep AI tools labeled as tools in internal systems, training, and SOPs, whatever guest-facing persona they wear.

• Name one accountable human, by role, for every AI-touched guest interaction.

• Don't let “the AI handled it” become an acceptable answer in a service-recovery review.

05. Open-Source Models Lower the Entry Price

SOURCE SIGNAL — “China's open-source bet.”

Chinese labs — DeepSeek, Z.ai, Moonshot, Alibaba's Qwen, MiniMax — are releasing frontier-class models as free, downloadable weights rather than metered APIs, partly to work around US chip export controls and partly to win developer goodwill. A recent study found that these open-weight models already account for roughly 17% of global AI model downloads.

WHAT IT MEANS FOR TRAVEL & HOSPITALITY

This is the clearest near-term win for independent hotels, small groups, and operators across Asia who were priced out of “Enterprise AI.” A genuinely capable model with no per-token bill changes the build-versus-buy math for anyone with in-house technical capacity.

• Independent and boutique properties: retire the assumption that AI is “for the big chains only” — it is now a capability and integration question more than a budget one.

• Regional associations such as PAIBA [Philippine AI Business Association] have a real opening to run shared open-source pilots that no single small operator could justify alone.

• Weigh the trade-offs honestly: data residency, model provenance, and support SLAs differ from proprietary vendors, and some labs' training methods remain under scrutiny.

06. Supercharged Scams Target the Booking Funnel

SOURCE SIGNAL — “Supercharged scams.”

Criminal use of AI has moved from crude spam to AI-generated phishing, deepfake voice and video, and malware that's harder to detect, deployed at a scale that overwhelms traditional defenses. Interpol has flagged scam centers across Southeast Asia switching locations and targets faster using cheap AI tools; Microsoft alone says it now screens over 100 trillion signals a day for this kind of activity.

WHAT IT MEANS FOR TRAVEL & HOSPITALITY

Hospitality sits in the blast radius by design: high-value transactions, third-party OTA and payment channels, guests primed to trust “your hotel” messaging, and front-line staff trained to be helpful rather than suspicious.

• Brief reservations, front-office, and finance teams specifically on AI-voice and AI-video impersonation, not only email phishing.

• Tighten verification steps for any payment, booking-change, or wire request that arrives with unusual urgency — even from a familiar voice or name.

• Treat this as a guest-trust issue as much as a security one: a scam wearing your brand's name damages the brand regardless of technical fault.

07. Humanoid Data and the Slow Robot

SOURCE SIGNAL — “Humanoid data” and “World models.”

Robotics companies are racing to collect the everyday-movement data — people microwaving food, wiping tables, opening doors — needed to train humanoid robots for real-world work, with over $6 billion invested in 2025 alone. But researchers openly admit today's robots still stumble over physics that simulations never quite capture, and nobody yet knows how many thousands of clips it would take to teach a robot to cook a single dinner reliably.

WHAT IT MEANS FOR TRAVEL & HOSPITALITY

Service robotics — delivery, back-of-house, limited food-and-beverage tasks — is real and improving. Still, the twelve months ahead are a data-collection and simulation-refinement phase, not a deployment wave. Treat any pitch for general-purpose humanoid staff accordingly.

• Fund robotics pilots for narrow, repeatable, low-consequence tasks — linen runs, fixed-route room-service delivery — rather than general-purpose “robot staff.”

• Ask any humanoid vendor what share of their training data came from real-world deployment versus simulation.

• Budget robotics as a multi-year capital position, not a feature for next year's opening.

08. The Trust Backlash and the Search Rewrite

SOURCE SIGNAL — “Resistance” and the AI Hype Index.

A visible anti-AI movement is building: a London march past the offices of OpenAI, Google DeepMind, and Meta; a cross-partisan “Pro-Human AI” declaration in the US; and polling suggesting roughly half of Americans are concerned about AI's growing role in daily life. At the same time, the discovery layer guests use to find you is being quietly rebuilt — Google is reportedly moving from a list of links toward an interactive, chatbot-led experience.

WHAT IT MEANS FOR TRAVEL & HOSPITALITY

Two things are true at once: guests are growing warier of being handled by AI, and guests are increasingly finding you through AI. Both point in the same direction — disclosure and differentiation, not concealment.

• State plainly, on-property and online, where AI is doing the work and where a person is — concealment erodes trust faster than admission does.

• Start treating Generative Engine Optimization — how AI answer engines describe and recommend your property — as seriously as SEO. It is the same discovery contest with a different referee.

• “The intelligence may be artificial, but the experience is human” only holds as a market position if it is demonstrably true at the front desk, not just on the website.

The 12-Month Roadmap

Three horizons, each building on the last. None requires a technology bet that doesn't already exist today; all require an owner.

MONTHS 1–3 (JUL–OCT 2026) — STABILIZE & DISCLOSE

Close the most exposed doors before opening any new ones. Brief every guest-facing team on AI-enabled scam patterns, audit vendor contracts for hidden compute-cost exposure, and put a plain-language AI disclosure on the website and at check-in. Nothing in this horizon requires a new capability — it requires attention.

MONTHS 4–8 (NOV 2026–MAR 2027) — ORCHESTRATE, DON'T OUTSOURCE

This is when agent orchestration becomes real enough to pilot, and popular enough to be pitched at every trade show. Run one bounded pilot, name a human owner for it, and resist any vendor language that describes the system as a new “team member.” Treat GEO as a live workstream, not a research topic.

MONTHS 9–12 (APR–JUL 2027) — DIFFERENTIATE ON TRUST

By this point, open-source and orchestrated AI will be table stakes for competitors who moved early. The differentiator left on the table is trust — properties that can show, not just say, where AI ends, and a person begins will be the ones the backlash-weary guest chooses. This is also when the robotics pilots from month one either graduate to wider rollout or get quietly retired. Decide the criteria for that now, not then.

Pertlink Bottom Line

None of the eight signals above is about hospitality. That is rather the point. AI's real constraints are being discovered in chip fabs, power grids, and cybercrime units months before they surface in a PMS release note — which means operators reading the engineering trade press right now have a longer runway than the ones waiting for the vendor webinar.

Build big, build small, or not at all: over the next twelve months, the wrong answer is the one made by default.

The intelligence may be artificial. But the experience is human — and for the next year, that will be the most expensive sentence in hospitality to get wrong.

About this paper

Synthesized from MIT Technology Review, Vol. 129 No. 4 (July/August 2026), read alongside Pertlink's ongoing research into hospitality AI economics, workforce design, and guest trust.

Made with the help of various AI tools, but with a HITL

Media Contact

Terence Ronson

Founder and Managing Director [email protected]