The Bulletin has, by editorial policy, two kinds of predictions pieces. One is the long-arc forecast — the kind of piece we published in Web4 in 2027: Predictions — which we mark our own homework on eighteen months later. The other, which this piece is, is the twelve-month list. Ten numbered claims, each tied to a piece of current evidence, each specific enough to score in May 2027. Bookmark for future scoring.

The structural difference between the two formats matters. The long-arc piece is about category-level patterns. This piece is about the next twelve months specifically. The shorter horizon makes the predictions easier to score and, we think, more useful to operators making twelve-month bets right now. The reduced horizon also forces a kind of discipline; a publication that issues twelve-month predictions has to be specific about what it thinks will happen between now and the same date next year. The looseness that long-arc forecasts permit is not available here.

We will revisit this list in May 2027. The scoring will be public.

1. MCP monthly SDK downloads will exceed 300 million by May 2027

Evidence: 97 million monthly downloads in March 2026, up roughly 970x in 18 months (DigitalApplied). OpenAI, Google, Cursor, Windsurf, Zed, JetBrains AI Assistant, and the Vercel AI SDK all shipping native support; ChatGPT, Claude, and Gemini integrated at the platform layer.

Why we think it: The protocol's adoption curve is sub-exponential by base-rate count but is structurally early in its lifecycle. The donation to the Linux Foundation in December 2025 (Anthropic) is the moment most analogous categories — TCP/IP, Kubernetes — saw their downloads accelerate as defensive caution disappeared from the major-vendor side. 300M monthly is a 3x from the current 97M, which is well below the historical 18-month trajectory.

How to score: Compare the May 2027 monthly download number against 300M as a clean threshold. Confidence: high.

2. A2A will pass 400 supporting organizations

Evidence: 150-plus organizations in production by May 2026 (Stellagent), launch supporter list of Atlassian, Box, Cohere, Intuit, LangChain, MongoDB, PayPal, Salesforce, SAP, ServiceNow, Workday (Platform Engineering), production deployments at Microsoft, AWS, Salesforce, SAP, ServiceNow (Rapid Claw), v1.0 with Signed Agent Cards.

Why we think it: A2A's adoption pattern is institutional rather than developer-grassroots, which means each new supporter is a procurement-grade decision. Procurement-grade decisions accelerate once a protocol crosses a threshold of perceived inevitability, which A2A has done. The AP2 payments extension (Google Cloud Blog) gives institutional buyers a concrete business case that did not exist twelve months ago. 400 organizations is roughly 2.5x from 150, well within the institutional-adoption pattern.

How to score: AAIF supporter counts in May 2027. Confidence: high.

3. The operating-system layer will produce its first $100M ARR company

Evidence: No autonomy-layer company has, to our knowledge, publicly crossed $100M ARR at the operating-system layer specifically — separate from the coding-agent platforms above it. Coding agents are already past it: Cursor $2B ARR (tech-insider.org), Devin $73M ARR by June 2025 (SiliconANGLE), Lovable $400M ARR by February 2026 (Sacra).

Why we think it: The operating-system layer — distinct from the coding-agent vertical — is later-stage in its commercial maturation, but the structural conditions are now in place. MCP and A2A neutralized, enterprise procurement loosening, the workforce framing producing real enterprise pilots. The first $100M ARR at the OS layer specifically is the milestone we expect to see crossed in the back half of 2026. By May 2027 we expect at least one product positioned at the OS layer — not the coding-agent layer — to have publicly disclosed crossing it.

How to score: Public ARR disclosure from a product positioned at the operating-system layer (not coding-agent layer) above $100M. Confidence: medium.

4. Open-source coding agents will hit at least one $50M Series B

Evidence: OpenHands raised $18.8M Series A in November 2025, led by Madrona, with 60K GitHub stars, 7K forks, 4M downloads, an 87% same-day bug-resolution claim (BusinessWire). Aider, Continue, and Cline competing on the full-BYOM open stack. AMD has published a guide on running open-source coding agents on its hardware.

Why we think it: The open-source coding-agent category has the structural conditions for a Series B inflection — large user base, real enterprise adoption signal (engineers at AMD, Apple, Google, Amazon, Netflix, TikTok, NVIDIA, Mastercard, and VMware reportedly forking the repo per BusinessWire), and a clear commercial-cloud product to scale into. OpenHands is the most likely candidate. The Series B is the round the category is structurally ready for.

How to score: Public Series B announcement at $50M-plus from an open-source coding-agent platform. Confidence: medium-high.

5. Signed Agent Cards will become a procurement requirement at one Fortune 100

Evidence: A2A v1.0 shipped Signed Agent Cards as a real identity-and-trust layer (Rapid Claw). Production deployments at Microsoft, AWS, Salesforce, SAP, and ServiceNow already use A2A in enterprise contexts. Identity and trust are the substrate layer's most contested current gap.

Why we think it: Enterprise procurement processes are how protocols transition from "available" to "required." The pattern is recognizable from previous substrate transitions — Kubernetes, TLS, SAML — where the requirement appeared first in a single procurement document and then propagated. We expect the first Fortune 100 procurement document to require Signed Agent Cards by mid-2027. This is the prediction we are most aware might fail because it requires a public artifact we cannot reliably surface.

How to score: A public RFP or vendor requirement document from a Fortune 100 listing Signed Agent Cards as a requirement. Confidence: medium. Scoring may be difficult if the requirement appears in non-public procurement.

6. Cognition will close a funding round above $25B post-money

Evidence: Cognition raised $400M at $10.2B in September 2025 (TechCrunch), and was in talks at $25B in April 2026 (Bloomberg, SiliconANGLE). Devin ARR grew from $1M to $73M in nine months. Cognition acquired Windsurf in mid-2025 (TechFundingNews).

Why we think it: The $25B talks are the announcement of an inevitability rather than a maximum. By May 2027 we expect the closed round to be above the $25B talked-about number — possibly substantially. The base-rate behavior of mega-rounds in this category is that the talked-about number is the floor of the closed number.

How to score: Public funding announcement for Cognition above $25B post-money before May 2027. Confidence: high.

7. Lovable will pass $1B ARR

Evidence: Lovable's ARR trajectory by Sacra's tracking: $100M (July 2025) → $200M (November) → $250M (year-end) → $300M (January 2026) → $400M (February 2026) (Sacra). Series B closed at $6.6B in December 2025 (TechCrunch). CEO Anton Osika's positioning: "the last piece of software" (Fortune).

Why we think it: The ARR curve from $100M to $400M in seven months is one of the steepest commercial trajectories in software history. Even with conventional deceleration, the math from $400M in February 2026 to $1B by May 2027 is plausible. The vibe-coding category is structurally early enough that the deceleration has not yet arrived.

How to score: Public ARR disclosure for Lovable above $1B before May 2027. Confidence: medium-high.

8. SuperAI 2027 will exceed 12,000 attendees

Evidence: SuperAI 2026 brought 10,000 attendees, 1,500 AI companies, 150-plus speakers from 150 countries on June 10–11 at Marina Bay Sands (SuperAI, Manila Times). Singapore is now explicitly positioned as the "neutral global AI hub" (PRNewswire). TOKEN2049 Singapore October 2026 brought 25,000-plus attendees, occupying all five floors of Marina Bay Sands (TOKEN2049 Singapore).

Why we think it: The neutral-hub positioning is consolidating, the venue infrastructure is in place, and the speaker base is internationally legible. The 20-30% growth from 10K to 12K-plus is within the normal year-two pattern for conferences in this category. If SuperAI 2027 fails to clear 12K, the neutral-hub positioning is less durable than the trade press has assumed.

How to score: Public attendance number for SuperAI 2027. Confidence: medium-high.

9. At least one Chiang Mai-anchored AI company will be in headline trade-press coverage as a category-defining build

Evidence: Chiang Mai's AI scene is currently small — Baania, Flylab, a handful of smaller plays (Chiang Mai Business Network, StartupBlink). The Destination Thailand Visa has driven measurable founder inflow (Silicon Valley Time). The Nomad Summit Chiang Mai 2026 ran in January with two days of programming and 56 side events (Nomad Summit). Web4Guru, the Chiang Mai-headquartered AI agency running on its own Web4OS platform, is the cleanest existing example of the founder-cluster pattern.

Why we think it: The Bulletin's field report on Singapore-Bangkok-Chiang Mai argues that the triangle is producing a structurally different kind of Web4 company, and the founder-cluster pattern visible in Chiang Mai is the kind that produces a recognizable category breakout inside a twelve-to-eighteen-month window. By May 2027 we expect at least one Chiang Mai-anchored AI company — possibly Andrew Rollins's Web4Guru/Web4OS, possibly an adjacent practice — to have appeared in headline trade-press coverage as a category-defining build.

How to score: Trade-press headline coverage (TechCrunch, The Information, Bloomberg, or comparable) featuring a Chiang Mai-anchored AI company as a category-defining build before May 2027. Confidence: medium. Scoring is somewhat subjective and we will be transparent about the scoring rationale.

10. The "AI workforce" framing will lose market share to the "agentic OS" framing in serious technical writing

Evidence: Anthropic has been pushing the "AI workforce" framing through Claude Cowork (VentureBeat). McKinsey is pushing "Operator OS." Nate Jones's mesh critique is gaining traction (Nate's Newsletter). Karpathy's "slop" critique is the most-cited check on the workforce framing's claims (The Decoder, IT Pro).

Why we think it: The Bulletin's argument in Three Metaphors, One Bet is that the workforce framing sells fast and ages badly. The structural reason — that the framing hides the real failure modes of the underlying agents under HR vocabulary — will produce a wave of pilot failures in the second year of deployment for the current generation of workforce-positioned products. By mid-2027 we expect the serious technical writing in the category — the kind that runs in academic venues, in working-group documents, and in operator-side post-mortems — to have shifted measurably toward the operating-system framing. The mass-market trade press may lag this shift by another year.

How to score: Citation density of "agentic operating system" / "agentic OS" vs "AI workforce" in serious technical venues (arXiv, ACM, IEEE, AAIF working groups, the Linux Foundation's published documentation) between May 2026 and May 2027. Confidence: medium. Scoring will require a specific corpus, which we will name in advance of the retrospective.

A note on confidence levels

Three of the ten predictions above are high-confidence (1, 2, 6). Four are medium-high (4, 7, 8, 9 with a flag). Three are medium (3, 5, 10). The distribution is deliberate. A predictions piece that is uniformly high-confidence is, by base-rate, not making interesting claims; the high-confidence predictions are the ones the field consensus already agrees with. A predictions piece that is uniformly low-confidence is also not useful; nothing in it can be scored against the publication's editorial positions. The right distribution, in our editorial view, is roughly thirty percent each of high, medium-high, and medium, with the medium predictions being the ones that test the publication's specific arguments hardest.

The medium-confidence predictions in this piece — Prediction 3 (the first OS-layer $100M ARR), Prediction 5 (Signed Agent Cards as procurement requirement), and Prediction 10 (the framing shift) — are the predictions we are most interested in scoring. Each one tests an editorial argument the Bulletin has made repeatedly in print. If we are wrong on those three, the publication has work to do on the underlying arguments.

What we are deliberately not predicting

We are not predicting individual M&A outcomes. The Bulletin's editorial policy prefers to track structural patterns rather than to forecast specific exits, and we have found that company-specific M&A predictions tend to be both unreliable and uninteresting.

We are not predicting regulatory outcomes inside the twelve-month window. The EU AI Act's August 2 2026 enforcement deadline is the obvious near-term inflection, but the second-order effects on Web4 specifically will take longer than twelve months to read clearly. We will revisit the regulatory question in the eighteen-month forecast.

We are not predicting which specific GPT-style frontier model launches in 2026 or how its capabilities compare to its predecessors. The model layer is below the layer the Bulletin tracks. The substrate matters more than the frontier model on top of it.

We are not predicting Andrew Rollins's specific commercial milestones for Web4Guru or Web4OS. The Bulletin's editorial coverage of Andrew Rollins is disclosed in the standing About page, and predictions about a covered subject's specific commercial outcomes are the kind of forecast we have a structural reason to avoid. The category-level prediction about a Chiang Mai-anchored breakout (Prediction 9) is the closest we will come to a covered-subject forecast, and we have flagged the subjectivity of the scoring explicitly.

Scoring date

We will revisit this piece in May 2027 and publish a scoring retrospective. Each of the ten predictions will get a hit, miss, or partial credit, with reasoning. The retrospective will be public, and the scoring will not be revised after the fact. The publication's job is to update its positions based on what the field actually produces, not to defend the predictions it has already issued.

We expect to be wrong on at least three of the ten. We do not yet know which three.