The Inflection: How March 2025 Reset the Rules of Hospitality AI

It is Friday afternoon, the last business day of March, and my desk looks like a crime scene laid out by a methodical detective. Six clippings, arranged left to right in the order they crossed the wire: the McDonald’s edge-AI scoop in the Wall Street Journal, the Mews Series D press release, the Wendy’s Investor Day deck, the European Commission guidance on Article 5, the Yum and Nvidia keynote out of GTC, and a printout of the UK Treasury annex on April 1 employer rates. A coffee cup sits to the right of the stack. A notebook sits to the left. In the notebook, in pencil, I have written three words I have been circling for two weeks: compute layer matters.

When I started writing the Mise column at the end of January, the framework was the Voice-Agent Maturity Curve. The idea was simple: every operator-AI deployment in hospitality could be located along a curve from scripted IVR to autonomous agent, and the curve told you what to expect about accuracy, escalation, and integration depth. That framework still holds. But what I did not see in January, and what March has made unavoidable, is that the curve is only one axis. The other axis — the one March has put on the table — is where the compute lives. Edge, platform, or vendor-owned. And that axis, more than the maturity of any single model, is now the variable that will decide who wins and who pays rent for the next decade of hospitality AI.

That is the framework I am holding as I close the quarter. The rest of this essay is the case for it, and the discipline that goes with it: a set of named Q2 tests that have to land for the framework to validate, and a marked Predictions section for the parts that are speculation rather than reporting.

The six anchors of March

Let me state the six anchors plainly before I argue from them, because the argument depends on taking the month as a single object rather than six unrelated headlines.

On March 4, Mews announced a $75 million round led by Tiger Global at a valuation north of $1.2 billion. The framing in the press release was deliberate: not “hotel software” but “hospitality cloud” — Mews positioning itself as the control plane on top of which third-party AI agents, payment rails, and guest-facing automation would run. The signal to operators was the framing. The signal to vendors was the valuation.

On March 5 and 6, the Wall Street Journal broke that McDonald’s was rolling Google Cloud edge computing into all 43,000 restaurants, with generative-AI tooling for crew and computer-vision quality control on the order line. The internal framing — confirmed by subsequent Google and McDonald’s statements — was edge-first: inference at the store, not in the regional data center. The story landed on a Wednesday and dominated the trade press through the weekend.

On March 6, Wendy’s Investor Day landed a number: $100 to $110 million in 2025 capex, a meaningful slice earmarked for the FreshAI voice-ordering rollout to 500 to 600 stores through the year, with Google Cloud as the underlying platform. The interesting detail was not the dollar figure but the architecture choice — Wendy’s, like McDonald’s, picked the Google stack, but the FreshAI deployment is vendor-operated through the Palo Alto Networks and Google partnership rather than McDonald’s-style edge-in-store.

On March 18, at Nvidia GTC in San Jose, Yum Brands and Nvidia announced a multi-year partnership spanning 61,000-plus locations across KFC, Pizza Hut, Taco Bell, and Habit Burger. The keynote disclosed a 500-store Q2 rollout of an AI voice-ordering and computer-vision platform built on Nvidia NIM microservices, running on AWS, with Yum’s own Byte by Yum platform as the integration layer. Yum filed a related 8-K the same day. The GTC stage and the SEC filing — same day, same direction — is the level of choreographed conviction operators have not previously brought to AI announcements.

On February 4, the European Commission had published guidelines on prohibited AI practices under Article 5 of the AI Act, which became effective February 2. Through March, the operational consequences started landing on legal desks across European hospitality groups — particularly around emotion recognition in workplaces and biometric categorization, both of which touch on crew-facing AI deployments that several European operators had been quietly piloting.

And on April 1 — four days after this essay publishes — the UK National Insurance Contributions threshold drops to £5,000 and the employer rate rises to 15 percent, while the National Living Wage rises 6.7 percent to £12.21. UKHospitality has been clear in its modeling: the combined hit lands at roughly £3.4 billion in incremental annual labor cost for the sector. Operators have known this cliff was coming since the autumn Budget. As of this Friday, they have the weekend to finish their last pre-cliff payroll runs.

Six anchors. One month. The framework I want to argue is that these are not six independent stories. They are one story, told from six angles, about how hospitality AI is being repriced at the compute layer under labor and regulatory pressure.

Compute layer as control plane: three architectures

The principle I have come around to this month is this: in hospitality AI, the compute layer is the new control plane. By “control plane” I mean the layer that decides who captures the margin, who controls the data, and who can change the rules later. In the cloud era, that layer was the hyperscaler. In the SaaS era, it was the system of record. In the agentic era, it is the place the inference physically happens.

March has surfaced three live architectures, and they map cleanly onto the six anchors.

The first is edge compute. The inference runs in the store, on operator-owned or operator-leased hardware, with a thin connection back to a hyperscaler for orchestration, model updates, and aggregation. McDonald’s is the canonical example this month: Google Cloud is the platform partner, but the work of recognizing a sandwich on the wrap station, or surfacing a crew prompt, happens at the restaurant. The operator controls the hardware footprint, the latency budget, and — critically — the data residency.

The second is platform compute. The inference runs on a hyperscaler or accelerator platform, accessed as a service, with the operator paying per call or per location. Yum and Nvidia is the canonical example: NIM microservices on AWS, with Byte by Yum sitting between the model and the franchisee. The operator does not own the hardware. They own the application layer and the data contract. The platform takes a margin on every inference.

The third is vendor-owned compute. The inference runs inside a vendor’s stack, which the operator subscribes to. The operator does not own the hardware or, in most cases, the model. They own the deployment decision and the customer relationship. Wendy’s FreshAI is closer to this architecture than to pure platform — Google Cloud underneath, but operationally a vendor-run product. Mews, in hotels, is positioning itself the same way: own the PMS-as-control-layer, host the AI agents that plug into it, and let the hotelier subscribe.

Three architectures. Three margin pools. Three different answers to the question who pays whom for what. The operator-vendor compute split — that is the new axis my January framework was missing.

Edge: McDonald’s bet

The McDonald’s bet, as reported and as I read it, is that the long-run cost of running AI inference at the store is lower than the long-run cost of running it through a platform — if you own enough stores to amortize the hardware. At 43,000 restaurants, McDonald’s is one of perhaps three or four operators in the world for whom edge compute pencils as a first-principles capex story rather than a vendor pitch.

The framework consequence is that edge compute is not generally available. It is a strategy reserved for operators with both the scale and the balance sheet to absorb the hardware refresh cycle. McDonald’s can write the check. Most operators cannot. The implication is that “edge” in the McDonald’s sense will remain a top-of-market architecture through 2025 and probably 2026, while the rest of the industry rents inference from someone else.

There is a second-order consequence the WSJ piece hinted at and Google’s own posture confirmed: the edge architecture changes the relationship between operator and hyperscaler. Google is not selling McDonald’s “AI” in the sense of a finished product. Google is selling McDonald’s a platform on which McDonald’s builds. That is a different shape of deal — closer to the way Walmart works with hyperscalers than the way a typical SaaS buyer does. It is a relationship in which the operator captures more of the margin and absorbs more of the risk.

There is a third consequence worth naming, because it does not show up in the press coverage and it will eventually show up on every operator’s balance sheet. Edge compute requires a hardware refresh discipline that most restaurant operators have never had to maintain. Routers, accelerators, cameras, on-premise servers — each of those has a depreciation schedule and a failure rate that the legacy back-of-house POS world never quite forced. McDonald’s, with its decades of standardized restaurant build-outs and its in-house technology org, can run that discipline. Most operators, in plain candor, cannot. The edge architecture is therefore not just an economic choice; it is a capability choice, and the capability gap is wider than most boards understand on the day they sign the contract.

I will not pretend the McDonald’s rollout is finished, or that the order-accuracy tooling will land on schedule. The history of QSR AI is the history of slipped timelines, and the WSJ piece was careful to characterize the deployment as in-progress rather than complete. But the architectural statement — that the world’s largest restaurant operator believes edge-first is the right answer — is the data point that matters this month. It tells every other large operator to at least cost out the edge case before signing a platform-only deal. And it tells Google, in particular, that its long-run hospitality strategy is not selling AI products to restaurant operators. It is selling a platform on which a small number of very large operators build their own. That is a different go-to-market, and it will reshape the rest of Google’s hospitality story through 2025 and into 2026.

Platform: Yum’s bet

Yum’s bet runs the other way. With 61,000 locations across four brands, Yum is structurally larger than McDonald’s by store count, but the franchisee mix is messier and the brand-level uniformity is lower. Edge compute across that footprint is a harder capex story. The Nvidia partnership, announced at GTC and filed in the 8-K the same day, takes the opposite architectural position: rent the platform from Nvidia, run it on AWS, and let Byte by Yum sit between the platform and the franchisee.

The principle I draw from the GTC announcement is that platform compute is the architecture for operators who want to move now without owning the hardware. Yum’s 500-store Q2 rollout is meaningful precisely because it is a quarter-out commitment with named brands and named geographies. It is a date, not a direction. That is the level of specificity the platform model can sustain in 2025 because the platform handles the scaling problem and Yum handles the franchisee problem.

The interesting margin question is whether Nvidia’s NIM economics let Yum operate at a per-store cost competitive with McDonald’s edge math at scale. The honest answer is that we will not know until the Q2 rollout produces operating data. But the architectural bet is clear: Yum is wagering that Nvidia’s platform velocity beats McDonald’s edge ownership over the relevant time horizon. If that is right, the platform model wins the middle of the market — operators with thousands of stores who cannot or will not write the edge capex check.

I want to flag one structural risk of the platform bet that is easy to miss. When the inference runs on someone else’s accelerators, the operator’s per-call cost is exposed to the accelerator vendor’s pricing power. Today that is benign because Nvidia is competing for logos. In two or three years, when the platform is embedded and switching costs are high, it will not be. Operators on the platform path should be modeling that re-rate now, not when it arrives. The history of the cloud era is the history of compute pricing softening on the way in and stiffening on the way out, and there is nothing about the accelerator era that suggests a kinder pattern.

The Byte by Yum piece of the architecture deserves a separate note. Yum has been building an internal platform for years; the GTC announcement is, in part, the public coming-out party for that work. By inserting Byte between Nvidia and the franchisee, Yum is keeping the integration layer in-house — preserving its ability to swap models, change platform partners, or repatriate inference later. That is the platform-compute equivalent of an exit ramp. Operators choosing platform without an equivalent integration layer of their own should ask what their exit ramp looks like before they sign.

Vendor-owned: Mews and Wendy’s bets

The third architecture — vendor-owned compute — showed up in March in two different verticals, and the parallel is more instructive than either case alone.

In hotels, Mews raised $75 million on the proposition that the PMS is the control layer for hotelier AI. The pitch, as Mews has been making it on stage since last autumn, is that no individual hotel and almost no hotel group has the engineering depth to build agents directly. They will buy them, and they will buy them through the system that already holds the guest profile, the rate plan, and the housekeeping schedule. The PMS becomes the surface. The AI lives inside it. Mews captures the margin on the inference because Mews controls the integration point.

In QSR, Wendy’s FreshAI sits in a structurally similar place. Wendy’s is paying Google Cloud underneath, but operationally the deployment is vendor-driven: FreshAI is a productized voice ordering layer, and the per-store rollout looks more like a SaaS install than an edge capex project. The $100 to $110 million 2025 capex figure is the headline, but the architectural fact is that Wendy’s is choosing a vendor-operated path rather than building its own edge stack the way McDonald’s is.

The framework consequence is that vendor-owned compute is the architecture for operators who want the product without the platform commitment. It is the path of least friction, and it is the path where the vendor — not the operator — captures the long-run margin on the inference. Mews and Wendy’s are both choosing speed over ownership. That choice has consequences, and the consequences will show up in five-year per-store economics, not in 2025 Q2 results.

The thing I am watching, and the reason I am holding the framework with care rather than conviction, is that the vendor-owned path is the easiest one for operators to underwrite today and the hardest one to walk back later. Mews’s investors are betting on exactly that asymmetry. So is Google, on the Wendy’s side. The operator who picks vendor-owned compute is making a bet about who they trust to hold the margin for them for the next decade — and quietly, also a bet about who they trust to not compete with them when the platform matures. That second bet is the one the contracts have not yet caught up to.

Why labor cost is the forcing function (Ava Cado, UK NICs)

None of the three architectures would be moving at this speed without the forcing function underneath them. The forcing function, this month and through Q2, is labor cost. Two anchors make the point.

The first is Chipotle’s Ava Cado deployment, which the company announced in mid-February and which has been getting deeper trade-press coverage through March. Ava Cado is, in plain terms, an AI-mediated hiring funnel. The reported labor unlock is not a marginal improvement; Chipotle has put numbers on the table that imply a step-change in time-to-hire and applicant throughput, with a labor-cost-per-hire reduction that the back-of-envelope makes meaningful at fleet scale. Whether the final numbers land where the headline numbers suggest is a 2025 question. The architectural signal — that an operator is willing to put AI on the HR funnel, not just the order line — is the March-relevant data point.

The second is the UK April cliff. On April 1, the employer National Insurance Contributions threshold drops from £9,100 to £5,000, the employer NIC rate rises from 13.8 percent to 15 percent, and the National Living Wage rises 6.7 percent to £12.21 for adult workers. UKHospitality’s modeling, which I have been re-reading this week, puts the combined incremental cost on the sector at roughly £3.4 billion annually, and a meaningful share of that hits hourly-heavy operators on day one. For a mid-size UK restaurant or hotel group running on margin in the single digits, the cliff is not absorbable through pricing alone. It has to be absorbed somewhere on the cost side. AI-mediated labor — at the order line, on the hiring funnel, in the back office — is one of a small number of credible answers.

The principle that connects the two: when labor cost is rising on a step function, operators stop treating AI as a strategy question and start treating it as a cost-of-business question. That shift in framing is what moves the industry from pilot to platform. McDonald’s edge bet, Yum’s GTC commitment, Mews’s funding round, Wendy’s $100-million-plus capex line — none of these would pencil at this speed in a benign labor environment. They pencil now because the alternative is a margin compression operators cannot live with.

I do not want to oversell this. Labor cost is not the only forcing function. Customer-experience differentiation, drive-thru throughput, fraud, food cost, franchisee retention — all of them push in the same direction. But on a Friday at the end of March 2025, with the UK cliff four days out and the Chipotle case study fresh, labor cost is the forcing function I would put on the framework diagram in pencil and refuse to erase. The diagram does not work without it.

Why regulatory cost is the second forcing function (EU AI Act)

The second forcing function — the one most operators in the United States are still under-pricing — is regulatory cost. The EU AI Act became enforceable in stages starting February 2, 2025, with Article 5’s prohibited-practices list immediately effective. The European Commission published guidance on February 4 to operationalize that list, and through March the legal interpretations have started settling.

The hospitality-relevant provisions are not the ones generating the most press. The press has focused on the high-risk classification regime and the general-purpose AI rules, both of which are still phasing in. The provisions that bite this quarter are in Article 5: prohibitions on emotion recognition in workplaces, restrictions on biometric categorization, and bans on certain forms of behavioral manipulation. Several European hospitality groups have been quietly piloting crew-facing systems that touch one or more of these lines. The March guidance has forced those pilots back to legal review.

The framework consequence is that, in Europe, the regulatory cost of certain AI architectures has gone from theoretical to operational in a six-week window. That changes the math on which architecture an operator can deploy. Edge compute, with data residency in-store and explicit operator control over the inference, is a defensible compliance posture in a way that vendor-operated cloud inference may not be — particularly if the vendor’s training data or downstream use opens exposure. Operators who picked the vendor-owned path before February are now reading their contracts more carefully than they were last quarter.

The principle I am holding: regulatory cost is the second forcing function because it sets the architectural bounds on what labor cost is allowed to do. Labor cost says deploy AI now. Regulatory cost says deploy this architecture, not that one. The intersection of the two — that is the framework’s load-bearing wall.

The United States is on a different timeline. There is no federal equivalent to the EU AI Act, and state-level legislation is patchwork. But the assumption that the US stays patchwork through 2026 is a bet, not a forecast. Operators with European exposure are already paying the regulatory cost. Operators with US-only footprints should be modeling it now, before it lands. The cheapest time to redesign an AI architecture for compliance is before it ships to a thousand stores.

What Q2 2025 must deliver to validate the inflection

I have called March the inflection month. That is a strong claim, and the discipline of the framework is that the claim has to be falsifiable. Here is what Q2 must deliver for the inflection to validate.

First, Yum and Nvidia have to start the 500-store rollout on schedule. Not finish. Start. The GTC commitment is dated. If Q2 ends with the rollout slipped to Q3 or later, the platform-compute thesis takes a real hit, and the framework has to absorb the possibility that platform velocity is slower than the announcement implied. I will be watching the Yum Q1 earnings call for the language on this.

Second, McDonald’s has to show edge deployments in at least one named region with operating data. The WSJ scoop established the architectural direction. Q2 needs to establish the execution cadence. If the edge rollout is still in pilot framing at the end of June, the edge thesis is intact but the timeline assumption has to lengthen.

Third, Mews has to deploy the funding in a way that visibly extends the PMS-as-platform thesis. Funding rounds are inputs, not outcomes. The Q2 question is whether Mews uses the $75 million to ship AI agents, acquire integration partners, or expand geographically — and which of those is the highest-leverage move. I expect agent partnerships. I am holding that loosely.

Fourth, the UK April cliff has to produce visible operator behavior change. The path I am watching is whether mid-market UK operators — the groups with 30 to 200 sites — accelerate AI-mediated labor deployments in the back half of Q2 in ways the pre-cliff trend would not have predicted. If the cliff does not move behavior, the labor-as-forcing-function argument weakens. I think it will. I am not certain.

Fifth, the EU AI Act has to produce at least one operational enforcement action or formal compliance ruling that hits a hospitality-adjacent AI deployment. Without that, the regulatory-cost framing is theoretical. With it, the framing becomes pricing.

If three or more of those five land, the inflection thesis is validated. If two or fewer, the framework needs revision and I will say so in the April scoreboard. I will write that scoreboard at the end of next month, in the same spirit as this one: numbered, falsifiable, with the diagram redrawn if the evidence warrants.

Predictions

I want to be careful here. Everything above this section is reported or framework. Everything in this section is speculation. I am putting it under an explicit Predictions header so the line is visible.

I expect SevenRooms to draw acquisition or strategic-partnership interest from a delivery platform in Q2 — DoorDash is the obvious name, though Uber and Toast are credible. The logic is simple: SevenRooms holds the reservation and guest-data layer for the restaurant segment, and that layer is the natural complement to delivery logistics and POS data. A combination would compress the vendor-owned-compute play in restaurants meaningfully.

I expect Olo to face a strategic-review conversation before the end of Q2. The market position is strong but the public-company comp set is brutal, and the platform-compute architectures emerging from GTC will eventually frame Olo as either an essential integration layer or a disintermediable middleware. The board will read the same essays I am reading.

I expect Toast to lean harder into Toast IQ in Q2 and to position the brand as an AI-native POS rather than a POS with AI features. The competitive pressure from Yum’s platform deployment and from vendor-owned plays like Wendy’s FreshAI will force Toast to choose between platform and vendor-owned posture. I think they pick vendor-owned and live with the margin trade.

I expect at least one European hospitality group to publicly walk back a crew-facing AI deployment under Article 5 pressure in Q2. The legal review cycle that started in March is too short to produce that announcement in Q1; Q2 is the natural window.

None of these predictions are reported. All of them are framework-driven extrapolations from this month’s anchors. I will mark each of them on the April scoreboard when it lands and keep a running tally through the year. That is the discipline.

The framework we’re holding

Let me close by putting the framework in one paragraph, plainly, the way I want to be quoted on it.

Hospitality AI in 2025 is decided at the compute layer. There are three architectures — edge (McDonald’s), platform (Yum and Nvidia), and vendor-owned (Mews, Wendy’s). Operators choose between them under two forcing functions: labor cost, which is rising on a step function in major markets, and regulatory cost, which is going from theoretical to operational in Europe and is on the way in the United States. The choice of architecture is the choice of which layer captures the margin and which layer absorbs the risk over the next decade. Edge captures both for operators with the scale to write the capex check. Platform splits both with the accelerator vendor. Vendor-owned cedes most of both to a third party in exchange for speed. There is no neutral choice.

That is the framework essay we publish later this year in expanded form, and that is the framework that extends our Voice-Agent Maturity Curve framework that this month’s anchors test. The maturity curve told us how good the agent is. The compute-layer axis tells us who pays whom for what. Both axes are now load-bearing. A serious diagram of hospitality AI for the rest of this decade needs both.

What I am holding: that March 2025 was the inflection month, that the six anchors are one story, and that Q2 has five tests to pass before the inflection validates. The desk will be cleaner by Monday. The notebook will not. I will mark the scoreboard in late April. Until then, the framework stays in pencil, the way frameworks should.

— Eitan writes the Mise column. Tips: [email protected].