Frontier AI  |  Physical-World AGI  |  June 2026

Visual AGI: Why Machines That Read Floor Plans, Wiring Diagrams and Crop Maps Are the Next Trillion-Dollar AI Category, and Why They Still Think Like Toddlers

Artificial intelligence has spent the past three years learning how to talk. It has not yet learned how to see, not in the way that matters for the physical economy. That distinction sounds like a marketing line until you put a number on it. In January 2026, a 29-author research team published BabyVision, a benchmark built specifically to strip away language and test the visual primitives that human children master before they can speak: counting occluded objects, tracking motion, judging spatial relationships, recognising patterns. The best frontier model tested, Google's Gemini 3 Pro Preview, scored 49.7 against a human adult baseline of 94.1, a result the researchers describe as lagging behind six-year-old humans on tasks three-year-olds solve effortlessly. Every other proprietary model tested, including GPT-5.2 at 34.4 and Claude 4.5 Opus at 14.2, fell further behind.

That gap is not a curiosity. It is the central commercial bottleneck standing between artificial intelligence and the trillions of dollars now being committed to physically building the AI economy, the data centres, factories, farms and buildings that language models cannot help design because they cannot reliably tell you which wire connects to which terminal. This article forensically maps that gap using the verified benchmark data, names the companies racing to close it, traces the capital already moving, and ends with a scoring tool for separating genuine visual reasoning infrastructure from generative-AI window dressing wearing a new label.

The Toddler Problem: What the BabyVision Benchmark Actually Found

The reason a frontier model can write working code, pass a bar exam and solve PhD-level physics problems while struggling to count blocks in a photograph comes down to how today's vision-language models actually process an image. Elorian AI, the visual reasoning startup co-founded by 14-year Google DeepMind veteran Andrew Dai, frames the limitation precisely: today's models reason in a two-step process, first translating visual inputs into language, then performing text-based reasoning on that translation. The chain is fragile because many visual tasks cannot be precisely or concisely described in words in the first place. You cannot caption your way to understanding which beam a pipe needs to clear, because the constraint is geometric, not linguistic.

The BabyVision results make the scale of the resulting gap concrete. The benchmark's 388 questions span four categories, fine-grained discrimination, visual tracking, spatial perception and visual pattern recognition, deliberately minimising language and prior-knowledge demands so scores reflect visual reasoning rather than text-based inference dressed up as vision. One subtask, counting 3D blocks under occlusion and viewpoint variation, produced a best score of just 20.5% across every model tested. DN's AI Agent Economy Tracker has documented similar fragility patterns in agentic tool use; this is the same brittleness showing up at the perception layer instead of the planning layer.

This sits alongside a genuinely more encouraging signal that complicates a purely pessimistic reading. On ARC-AGI-2, a benchmark designed to test abstract, non-memorised visual pattern recognition, Google's Gemini 3 Pro scored 31.1% in November 2025, nearly double GPT-5.1's 17.6%. By February 2026, Gemini 3.1 Pro had pushed that verified score to 77.1%, more than doubling the prior generation's result in roughly three months. That is the honest, two-sided picture: foundational visual primitives (counting, occlusion, spatial tracking) remain stubbornly toddler-level, while abstract visual pattern recognition on curated puzzle-style benchmarks is improving at a pace that outstrips almost every other capability gain in the industry's history. Whether ARC-AGI-2 gains transfer to messy, real-world engineering diagrams, the actual commercial prize, is the open question the rest of this article is built around.

Why Data Centres Became the Most Urgent Visual Reasoning Use Case in the Economy

The reason visual reasoning has stopped being an academic curiosity and started attracting nine-figure strategic checks is that the physical economy is being rebuilt around AI infrastructure at a pace and capital intensity with no modern precedent, and every dollar of it runs through visual-spatial problems language models cannot touch. McKinsey's own published estimate for global data centre capital expenditure through 2030 has moved within the firm's own research in the past year, from an April 2025 estimate of $6.7 trillion (covering the full build-out including IT hardware) up toward the $7 trillion figure now commonly cited, with the firm's separate March 2026 analysis on industrial supply chains describing the same build-out as a $7 trillion race. Within that figure, capacity equipped specifically for AI processing loads accounts for roughly $5.2 trillion, against $1.5 trillion for traditional IT workloads, and McKinsey projects global data centre capacity will almost triple by 2030, from roughly 103 gigawatts today to around 219 to 220 gigawatts.

The reason this is a visual reasoning problem and not simply a financing problem is structural. A modern AI data centre, what NVIDIA now calls an "AI factory," is a visual-spatial puzzle with extreme financial consequences for every error: rack layouts, cable trays, power distribution paths, liquid cooling loops, backup generators, switchgear, fire suppression systems and maintenance access zones, all of which must reconcile with each other under real-world physical constraints before a single watt of compute comes online. As Elorian's Dai has put it, current models cannot reliably tell you what two things a wire is connected to, which is precisely the kind of question that matters when you are building a facility at gigawatt scale, where a single misrouted cable run or undetected thermal clash can cost months and tens of millions of dollars.

The industry's own infrastructure response confirms how seriously this is being taken at the engineering layer. NVIDIA's Omniverse DSX Blueprint, released alongside the Vera Rubin DSX AI Factory reference design, lets builders and operators model entire AI factories in physically accurate virtual environments before construction begins, unifying power, cooling and networking simulation into a single digital twin built on Universal Scene Description standards. The blueprint's industry support list reads like a roll call of the entire data centre supply chain: PTC is integrating it into Windchill product lifecycle management to connect engineering data with real-time simulation; Procore is building a continuous digital thread across the construction lifecycle; Trane Technologies is using it to optimise thermal management for gigawatt-scale facilities; AVEVA, Schneider Electric, Siemens and Dassault Systèmes are contributing engineering and simulation software directly into the blueprint. None of this works without a model underneath it that can actually read the diagrams these systems generate and reconcile them against real-world photographic evidence of what got built.

The Benchmark Snapshot: Where Frontier Models Actually Stand on Visual Reasoning, June 2026

Read across the row, and the pattern is consistent: the same models that have closed most of the gap to human performance on language, code and structured mathematics remain well behind on tasks that require holding a coherent spatial or visual state, the exact capability that engineering, architecture, construction and agriculture actually need.

Who Is Actually Building This: The Named Companies Racing to Close the Gap

The clearest signal that visual reasoning has moved from research curiosity to investable category is who is writing the checks and who is leaving secure positions at the largest labs to chase it. Andrew Dai spent close to fourteen years at Google, including over a decade across Google Brain and DeepMind, before co-founding Elorian AI in Palo Alto with Yinfei Yang, a former Apple multimodal AI lead, and Seth Neel, a former Harvard assistant professor. The company raised $55 million in a round led by Striker Venture Partners, Menlo Ventures and Altimeter, with participation from NVIDIA and Google's own Jeff Dean, at a reported $300 million valuation. Dai's framing of the problem is blunt: he has said the artificial intelligence models at big labs have roughly the visual sophistication of a three-year-old, and the company's entire thesis rests on the BabyVision-style observation that perception precedes language in human cognitive development, while today's AI architecture inverted that order, building visual understanding as an afterthought bolted onto language models that were never designed to hold a spatial state.

Autodesk's response has been to build internally and to write the largest strategic check in the company's history simultaneously. The firm's neural CAD initiative, announced at Autodesk University in 2025, represents what the company describes as a complete reimagining of the parametric CAD engines that have remained largely unchanged for forty years. Unlike a general-purpose language model bolted onto existing software, neural CAD foundation models are trained directly on CAD geometry, learning to break apart and resynthesise the faces, edges and topology of professional design data. Autodesk's own senior VP of research, Mike Haley, has demonstrated the system generating multiple complete product design iterations from a single text prompt, and the company claims the technology could eventually automate 80 to 90% of routine design tasks. Separately, in February 2026, Autodesk invested $200 million in World Labs, a physical AI startup focused on 3D model generation, taking the startup's total raise to $1 billion and marking, in the company's own description, the largest single startup investment in Autodesk's history.

NVIDIA's bet: simulate the factory before you pour the concrete

NVIDIA's approach treats visual reasoning less as a standalone model category and more as an infrastructure layer underneath every other physical AI investment. The Omniverse DSX Blueprint, paired with the Vera Rubin DSX AI Factory reference design, allows data centre builders to model power, cooling and networking topology in a physically accurate digital twin before a single piece of equipment is installed on site, with partners across the entire supply chain, from Vertiv's prefabricated infrastructure to Schneider Electric's power distribution modelling, building directly into the blueprint rather than around it. NVIDIA has separately pushed Physical AI Data Factory tooling built on its Cosmos world models, designed to generate the synthetic visual training data that narrow industrial models will need, since real-world labelled imagery of, say, a thermal fault in a half-built gigawatt campus is inherently scarce.

Two Forensic Case Studies: Where the Capital Is Actually Landing

The data centre case: a $7 trillion build-out with a visual bottleneck at every layer

McKinsey's data centre demand model projects that global capacity will nearly triple from roughly 103 gigawatts today to between 200 and 220 gigawatts by 2030, requiring an estimated $6.7 trillion in cumulative capital expenditure, with roughly 70% of that demand driven directly by AI workloads rather than traditional IT. Separate McKinsey analysis on construction efficiency specifically estimates that reaching the full potential of modern data centre construction practices, many of which depend on better visual coordination between design files, site photography and as-built reality, could shave 10 to 20% off both delivery time and per-project capital spend, a saving McKinsey itself sizes at up to $250 billion against the narrower $1.7 trillion non-IT-hardware infrastructure spend through 2030. The build-out is so capital-intensive that hyperscalers Amazon, Alphabet, Microsoft and Meta alone are projected to spend roughly $725 billion on AI infrastructure collectively in 2026, while Goldman Sachs has documented a current US data centre capacity shortfall already exceeding 11 gigawatts, a gap expected to surpass 45 gigawatts by 2028 if construction cannot accelerate.

The venture case: physical AI absorbed a disproportionate share of a record funding quarter

Crunchbase data shows global venture investment hit an all-time quarterly record of roughly $300 billion in Q1 2026, more than 150% higher than the prior quarter, with frontier labs OpenAI, Anthropic and xAI alongside autonomous vehicle company Waymo absorbing $188 billion of that figure between them. Strip out the four largest megarounds, and a separate tracking effort identified 27 physical AI startups that each raised more than $50 million in the same quarter, collectively pulling in over $6.4 billion, with robotics absorbing roughly $4 billion of that total and seven companies closing Series A rounds above $200 million each, a round size and stage combination one analysis called a historically unusual pattern reflecting how capital-intensive physical AI development cycles have become relative to pure software startups.

Beyond Data Centres: Where Visual Reasoning Compounds Next

Data centres are the most urgent case because the capital intensity makes every visual reasoning error expensive in a way that is immediately measurable, but the underlying capability gap is identical across every domain built on diagrams rather than paragraphs. In construction more broadly, drone-based progress monitoring has moved from pilot programme to industry baseline: survey costs have fallen 60 to 80% as hardware costs have dropped 40 to 60% since 2020, and platforms processing that imagery against computer vision models can now flag safety risks at 85 to 95% accuracy in favourable conditions, identifying issues like missing personal protective equipment, uncovered excavations and scaffold irregularities automatically. McKinsey separately projects AI can lift construction productivity by up to 20%, cut costs by up to 15% and improve delivery times by up to 30%, though adoption remains genuinely early: by one estimate only around 12% of construction firms are using AI regularly today, which means the productivity case is still mostly a projection rather than an industry-wide verified outcome.

In architecture, Autodesk's neural CAD for buildings, now rolling into Forma, targets the specific constraint-aware problem that a glossy AI-generated rendering cannot solve: when an architect changes a building's shape, the system can instantly recompute internal walls, columns, structural grid lines and platform layouts to maintain spatial consistency, the kind of relational reasoning that BabyVision-style benchmarks show current general-purpose models still struggle with on basic occluded-object tasks. In agriculture, the same underlying capability, turning satellite imagery, drone footage and field photographs into structured, actionable decisions about irrigation stress, pest damage and yield variation, maps directly onto the crop health and remittance-adjacent financial inclusion themes that DN's Africa Stablecoin Dynamics coverage has tracked, since precision agriculture's data layer increasingly determines access to crop insurance and microfinance across smallholder farming regions where visual crop assessment has historically been manual and slow.

How to Actually Evaluate a "Visual AGI" Pitch, Given the Honest Complication Above

Do not take a "visual AGI" or "physical AI" label at face value; the term is being applied to everything from genuine CAD-reasoning foundation models to standard generative image tools with a rebrand, and the BabyVision and Pix2Fact results above confirm that even the strongest frontier general-purpose models remain far below human performance on the foundational visual primitives that real engineering and construction work depends on. Do treat the five forensic questions in the scoring tool above as a minimum filter before evaluating any vendor's claims: a company that cannot demonstrate diagram-level reasoning, editable structured output, native workflow integration, a compounding proprietary data advantage and at least pilot-stage measurable outcomes is, by definition, earlier-stage than its marketing implies, regardless of funding round size. Watch the concentration of capital specifically toward companies with credible domain-specific data moats, since Autodesk's own internal neural CAD investment alongside its external $200 million bet on World Labs signals that even the incumbents with the deepest proprietary CAD datasets in the industry do not believe a single approach will win outright.

For African crypto and fintech operators specifically, the visual reasoning wave is more directly relevant than it first appears: data centre buildouts increasingly determine sovereign compute capacity and AI infrastructure access across the continent, while precision agriculture's vision layer is becoming a quiet but material input into crop-backed lending and parametric insurance products that overlap with the financial inclusion use cases this publication tracks closely. See DN's Institutional Adoption Tracker for the broader infrastructure capital context this sits inside, and the DN's AI Agent Economy Tracker for how visual reasoning gaps compound with the agentic tool-use failures already documented in language-only systems.

What This Index Does Not Claim

The Bottom Line: The Capability Gap Is Real. So Is the Capital Racing to Close It.

The honest version of this story has two halves that most coverage collapses into one. The first half is that today's frontier AI models, for all their fluency in language, code and structured mathematics, remain forensically documented to perform at roughly toddler level on the basic visual primitives, counting under occlusion, tracking, spatial relation, that engineering, architecture, construction and agriculture actually run on. The second half is that the capital allocators closest to the physical economy, Autodesk, NVIDIA, the hyperscalers building $7 trillion worth of data centres, and a venture market that pushed $6.4 billion into physical AI startups in a single quarter, are not waiting for that gap to close before betting on who closes it. Visual reasoning will not arrive as a single dramatic breakthrough with an iPhone moment. It will arrive, as it already has in NVIDIA's Omniverse Blueprint partner list and Autodesk's neural CAD rollout, as a quiet productivity layer inside the industrial software the physical economy already runs on, narrow and high-value before it is ever general.

Embed grant: The DN Visual AGI Commercialisation Index may be reproduced with attribution to decentralised.news.
DN-INTERNAL links to resolve: DN Institutional Adoption Tracker, DN AI Agent Economy Tracker, DN Africa Stablecoin Dynamics.
Sources: arXiv 2601.06521 "BabyVision: Visual Reasoning Beyond Language" (Jan 2026), Vellum/DataCamp Gemini benchmark analysis (2026), McKinsey "The cost of compute: A $7 trillion race to scale data centers" (2025/2026), McKinsey "Scaling bigger, faster, cheaper data centers" (2025), NVIDIA Newsroom "Vera Rubin DSX AI Factory Reference Design" (2026), Bloomberg "Ex-Google DeepMind Researchers Debut Startup Called Elorian" (Apr 2026), Elorian.ai company manifesto, AEC Magazine "Autodesk shows its AI hand" (2025), Construction Dive "Autodesk invests $200M in AI startup" (Feb 2026), Crunchbase News Q1 2026 venture data, Foundevo "27 Physical AI Startups" (Mar 2026), Dan Cumberland Labs / DroneDeploy construction AI data (2026).
As of: June 2026. Benchmark scores, funding figures and capex projections reflect verified reporting as cited; figures from forecasting models (McKinsey capex, ARC-AGI/BabyVision benchmark scores) are subject to revision as new data is published. Not financial or investment advice.