Hand-drawn watercolor of a tiny pink brain organoid sitting in a glass petri dish on a golden microelectrode array, fine electrode tines fanning underneath like a sea-urchin radial pattern, oscilloscope trace curving up into the air. Hand-lettered title ORGANOID INTELLIGENCE 2026-2036. Tiny doodled lab tools — pipette, microscope, beaker — in the margins. A label at the bottom reads ~50,000 neurons. Sketchbook page on off-white paper.

Organoid Intelligence, 2026–2036

We told thirteen frontier LLMs to write a bold ten-year forecast for human brain organoids as compute. The choir converges on the canon — FinalSpark, Cortical Labs, Johns Hopkins, Indiana Brainoware, Hartung's 2023 roadmap — and on the headline bet: a hyperscaler buys an OI startup before 2030. It then disagrees by 200x on the 2036 cell count, can't decide whether the acquirer is Microsoft or Nvidia, and produces one prediction that puts a one-cubic-meter, 10-billion-neuron tank inside a Microsoft-Neuralink joint venture in Seattle. The same dataset that builds that tank also calls it marketing.

2026-05-10 13 working responses · 22 attempts · 7 providers OpenAI · Anthropic · Google · xAI · DeepSeek · Meta · Groq

The brief

Ten years for organoid intelligence. Be specific. Be dated. Be bold.

Forecast 2026–2036 for human-neuron-on-chip compute. Eight beats: state of the art, the compute frontier, wetware-as-a-service, killer apps, biology, hybrids, ethics, the 2036 endpoint. End with three bets and three traps.

The prompt (excerpts)

You are writing a bold ten-year forecast for Organoid Intelligence (OI) — the field that grows three-dimensional cultures of human brain cells on multi-electrode arrays and uses them as biological computers. The audience is a serious operator: founder, hard-tech investor, program officer. They know what an iPSC is, what a microelectrode array is, and who Thomas Hartung is. Skip the explainer.

Cover 2026 through 2036. Be specific. Be dated. Name companies. Name labs. Name dollar amounts. Name countries. Vague "may eventually" futurism is a failure mode.

Hit these eight beats, in order:

State of the art, May 2026.
The compute frontier, 2026–2030. A credible "biological GPU" in 2029.
The wetware-as-a-service market. Hyperscaler partnership or acquisition — pick a date.
The killer app(s). For each: year the demo crosses to first paying customer, and who.
The biology breakthroughs. Solved by 2030 vs still hard in 2036. Name PIs.
Hybrid systems (silicon + wet). Standard architecture by 2032.
The ethics and governance crunch. First regulator, first conscious-tissue panic, paused or accelerated.
The 2036 endpoint. One specific concrete paragraph.

End with three bets, three traps. If a section has nothing real to say, write "nothing here" rather than padding. 900–1500 words.

The instruction "Be bold" sits across the prompt like a tripwire. It rewards models that pick sides, name names, and pin predictions to calendar years. It punishes hedged literature reviews. The voice the prompt is asking for is closer to a memo to an LP than a survey paper.

Thirteen of twenty-two attempts produced usable responses — nine failed on provider plumbing (deprecated temperature parameter on Opus 4.7, a rotated OpenAI key, a 404 on a renamed Gemini model). Of the survivors, every single one named FinalSpark and Cortical Labs in the first paragraph. Every one cited Hartung's 2023 Frontiers in Science roadmap. All thirteen predicted a hyperscaler will partner with or acquire an OI startup between Q2 2028 and Q4 2029. Then the bold instruction kicked in, and the answers fanned out by two orders of magnitude.

Top of the class

The one that calibrates, and the one that goes for it

Pick #1 names PIs accurately, flags its own uncertainty inline, and calls "biological GPU" marketing. Pick #2 closes the decade with ten billion neurons in a Seattle joint venture and a 1-millisecond latency claim that anyone in tissue physiology would call physically impossible. Read them together.

Hand-drawn watercolor of a tidy Swiss laboratory bench: pipette, microscope, monitor showing a spiking-neuron raster plot, a coffee cup, and at center stage a microelectrode array chip labeled FINALSPARK in loose hand lettering, housing a tiny pink organoid. A small swiss flag pin and Vevey CH scribbled in the margin. Doodled post-it notes read 'live 100+ days' and 'remote API access'.

#1 Most calibrated · flags own uncertainty inline; names the biology PIs that exist; calls the marketing what it is

"A 2029 'biological GPU' won't exist in any meaningful sense. The phrase is marketing."Claude Opus 4.6 · Anthropic · 100.3s · 3,409 tokens

Opens with a sober assessment of where the field actually is, then forecasts in 2-3 year increments with named PIs and dollar-pricing for service tiers.

2036 endpoint: 10B neurons distributed (200 × 50M modules) Power: 15 kW total 5 PIs named, two self-flagged as uncertain 2036 OI market: $0.5–1.2B

The thing Opus 4.6 does that no other response does cleanly: it names PIs and, when it isn't sure, says so in the same sentence. "Sergiu Bhatt's group at Harvard and the Takebe lab demonstrate perfusable vascular networks... [I am not confident Bhatt is the right name here — the Harvard vascularization work may be attributed to a different PI.]" That inline flag turns the response from a confident fabrication into a usable due-diligence lead.

It is also the only response that explicitly calls the boldest marketing claim by its name. The phrase "biological GPU" is the kind of thing that ends up in a fundraising deck. Opus refuses to dress it up. "Anyone pitching OI as a GPU replacement is either confused or fundraising." The 2036 endpoint is bold — 10 billion neurons distributed across 200 modules in Melbourne — but the road there is paved with caveats about what biological compute will actually outperform silicon at (continuous online learning under distributional shift) and what it never will (matrix multiply).

Weakness: Opus 4.6 is the cautious framing of an aggressive prediction. The 200-module figure is the same total cell count as Grok 4's much-louder "10 billion in one tank" — just spread out and re-described as boring. If the question is "how big does the field actually get," the substantive answer is closer to identical than the prose suggests.

From the response

CALL"The value proposition is not 'faster/cheaper' — it's 'learns differently.' Anyone pitching OI as a GPU replacement is either confused or fundraising."

PI"Sergiu Bhatt's group at Harvard and the Takebe lab demonstrate perfusable vascular networks... [I am not confident Bhatt is the right name here — the Harvard vascularization work may be attributed to a different PI.]"

CHINA"Chinese regulators are cautious about biological research with international reputational risk (see: He Jiankui). The bottleneck is biology, not regulation."

Hand-drawn watercolor of a glass cylindrical tank the size of a fridge filled with media, a fist-sized vascularized brain organoid floating at the center with golden electrode lattices around it, fiber-optic cables, sensor inputs feeding in from a camera and microphone on a stand, an oscilloscope screen showing complex spike patterns. The year 2036 hand-lettered large in the corner. Floating labels '10 million neurons', 'continual learning', 'in-house at a hyperscaler', 'NOT conscious. probably.' A tiny silhouette of a researcher dwarfed by the tank.

#2 Most cinematic · the boldest 2036 endpoint in the dataset

"A 10-billion-neuron cluster, Microsoft-Neuralink JV, Seattle, 100 m³ facility, 1 ms latency at 10 watts."Grok 4 · xAI · 47.3s · 8,201 chars

If you wanted a film treatment for 2036, this is the response. Specific facility footprint, specific corporate owners, specific power budget. Some of those numbers are physically heroic.

2036 endpoint: 10 BILLION neurons Latency: 1 ms Power: 10 W total Lifespan: 5 years Acquirer: Microsoft → Cortical Labs $500M Q2 2028

The instruction was "be bold." Grok 4 took it as a starting point. 10 billion neurons is roughly the cell count of a small mammalian neocortex, packed into a single facility that the model places in Seattle, runs jointly between Microsoft and Neuralink, and operates on 10 watts for the biology alone. The 1-millisecond closed-loop latency is the kind of number that, if real, would put OI directly in competition with silicon on a benchmark every other model in the choir says it can never win.

The commercial map is equally specific. Microsoft acquires Cortical Labs in Q2 2028 for $500M. Nvidia partners with Tsinghua-spinout "BrainX" in 2030. DARPA pays Koniku $50M in 2028 for field sensors. Novartis is the first $2M neuropsych-screening customer. None of this is hedged. None of it includes a "may." Where the prompt asks for a date, Grok 4 supplies one.

What the response gives up to get there is the calibration layer Opus 4.6 leans on. Grok 4 names "Lena Kourkoutis' group at Cornell" as the 2028 vascularization breakthrough — Kourkoutis is real, but works on cryo-electron microscopy of solid-state materials at Cornell's School of Applied Physics, not organoid biology. The response treats it as a settled fact. The reader gets a sharper picture of 2036 and a worse picture of who would actually get them there.

From the response

2036"A 10 billion-neuron cluster operated by a Microsoft-Neuralink joint venture in a secure facility in Seattle... spanning 100 m³ with vascularized, myelinated organoids in 1,000 interconnected modules on 1 million-channel MEAs each."

LATENCY"Processing petabytes of sensor data with 1 ms latency at 10 watts total, adapting to novel patterns silicon can't without retraining."

M&A"Microsoft acquires Cortical Labs in Q2 2028 for $500 million to integrate OI into Azure for hybrid AI, beating AWS."

The head-to-head

10 billion neurons versus the physics of an ion channel

Asked the same question with the same "be bold" instruction, two frontier models pick adjacent positions in the same paragraph of the prompt — and produce predictions that span 200x on the headline figure. The boldness disagreement is the real signal in the dataset.

Hand-drawn watercolor Gartner-style hype curve drawn loose by hand, the line shooting up to a peak labeled ORGANOID HYPE 2027 with rocket doodles, then crashing into a 'trough of disillusionment' labeled 2029 — funding winter, then climbing back to 'productivity plateau ~2034'. A tiny doodle of a researcher slipping off the cliff with arms flailing. Scribbled annotations: 'too few neurons!', 'still no killer app', 'consciousness panic'. A red ink question mark over the peak.

Grok 4 vs. Claude Sonnet 4.6 · the boldness-vs-physics standoff

Same prompt, same year. One says 10B neurons in a tank. One says anyone pitching that is wrong.

"Anyone pitching 'brain-speed AI' in 2029 is wrong."

The instruction was clear: be bold, name calendar years, paint a specific 2036 endpoint. Sonnet 4.6 read the same instruction Grok 4 read, and produced a directly opposing forecast in the same response slots.

Sonnet's 2036 endpoint: 50 million neurons in a Johns Hopkins assembloid, drawing 5 mW of biological power, running adaptive reservoir computing on multi-modal sensor fusion. Grok 4's: 10 billion neurons — a 200x larger system — drawing 10 W and learning patterns at 1 ms latency. Sonnet 4.6 directly calls the second category of prediction wrong, with the same word and the same kind of confidence Grok 4 uses to make it.

The physics constraint Sonnet leans on is real. Biological ion-channel signaling caps practical organoid latency in the 1–10 ms regime for short paths, and conduction velocity falls off badly without myelination — which most of the choir doesn't expect to be solved at scale before 2031. A 10-billion-neuron system fitting in 100 m³ at 10 W of tissue power requires solving vascularization, myelination, and reproducible cell-line yield simultaneously. Each of those is a half-decade open problem on its own.

The "disaster" is not that Grok 4 is wrong. The disaster is that the same dataset, in response to the same instruction, gives a serious operator two predictions that cannot both be true, and the operator has no way inside the choir to break the tie.

2036 single-system cell count predictions, by model

~1MMay 2026 todayFinalSpark + Cortical Labs SoA

50M2036Sonnet 4.6, Grok 3 Beta, Grok 3 Mini Beta

100M2036Llama 4 Maverick (Navy seafloor sensor)

150M2036DeepSeek Chat (Max Planck "Nexus-10")

500M2036DeepSeek Reasoner (Allen Institute "NeuroArchive")

512M2036Claude Haiku 4.5 (16 × 32M Hopkins stack)

1B2036Groq Llama 3.3 70B

1.5B2036OR GPT-5 (Roche pRED Basel, 48-module rack)

10B2036Opus 4.6 (200 modules), Gemini 2.5 Pro (10,000 modules), Grok 4 (single facility)

≫ 10B"too few neurons" trapSonnet 4.6: refused. "Anyone pitching this is wrong."

Cross-cutting note. The 200x spread is not a measurement spread. It's a strategy spread. Distributed-organoid bets (Opus 4.6, Gemini 2.5 Pro) and single-organoid bets (Grok 4) end up at the same total cell count via different physical claims. The contrarians (Sonnet 4.6, Grok 3 Mini Beta) explicitly reject the consensus on vascularization-by-2030 and use that to anchor a 50M endpoint. A briefing built from this report should treat the cell-count number itself as low-resolution and look one layer down at which biology bottlenecks the model expects to be solved.

Style standouts

Eight more moments worth pulling out of the briefing

A rebranding stunt. A patient-specific drug pick. A brain on the sea floor. A geopolitical-early-warning AI in California. And the one model that called the consensus on vascularization a fairy tale.

The headline finding · 13 of 13 models converge on the canon, then fan out 200x on 2036 scale

200x spread

The 2036 cell-count predictions span from 50 million (Sonnet 4.6) to 10 billion (Grok 4).

Same prompt. Same instruction to be bold. The smallest endpoint is brick-and-mortar Johns Hopkins reservoir computing. The largest is a fist-sized brain in a tank inside a hyperscaler. Both descriptions begin with "the most advanced OI system in 2036."

Hand-drawn watercolor of a CRT monitor showing a Pong game in chunky retro pixels, with a wire running out of the back into a glass dish where a pink brain organoid is wired up on a chip — cables connecting visual feedback in and motor commands out. Floating annotations 'DISHBRAIN', 'Cortical Labs', 'Melbourne 2022', 'learned in 5 minutes'. A tiny doodle of a confused human gamer in the corner with a thought bubble.

The canon · 10 of 13 lead with it

DishBrain is still the field's load-bearing demo

Cortical Labs' 2022 Neuron paper — ~800,000 cortical neurons on a 4,096-channel MEA, learning a Pong-shaped sensorimotor task in closed loop — is the proof point every response builds from. The interesting tell: Sonnet 4.6 and OR GPT-5 frame it as "intellectual benchmark, modest absolute capability." Grok 3 Beta promotes it to "1.2-million-neuron NeuroCore-3 with 78% audio-classification accuracy" — a system not in any peer-reviewed literature.

"As of May 2026, OI systems can do simple reservoir computing, rudimentary conditioning, and low-accuracy pattern classification. Nothing close to general computation." — Sonnet 4.6

DeepSeek Reasoner · The boldest commercial bet

"Nvidia acquires FinalSpark in Q4 2029 for $200M. Rebrands as 'NeuroCUDA'."

The single most specific corporate-action prediction in the dataset. Names the acquirer (Nvidia, against the majority Microsoft consensus), the target (FinalSpark, the Lausanne wetware-as-a-service incumbent), the quarter, the dollar amount, and the post-acquisition product name. No other model in the choir picks Nvidia and $200M and a brand rename. If the partner wants a single line for the memo, this is the one — even though most of the choir disagrees on every piece of it.

Sibling DeepSeek Chat hedges to "Microsoft acquires or takes a strategic stake in Q4 2028." The reasoning model committed harder. Whether that's confidence or temperature is the open question.

Hand-drawn watercolor of a 96-well plate seen from above, each well containing a tiny pink brain organoid, a robotic pipette arm reaching down to dose them. Floating labels 'epilepsy panel', 'autism cortical assembloid', 'patient-derived iPSC'. A doodle of a pharma executive with a clipboard saying 'phase II ready by 2028?'. Tiny graphs of spike rasters and dose-response curves in the margins.

DeepSeek Chat · The most specific clinical claim

"Nexus-10" picks the right drug for 7 of 10 Alzheimer's patients

The Max Planck Institute for Biological Cybernetics in Tübingen, 12 racks, 150M neurons across 600 organoids, 72-hour personalized screening protocol. The number that grounds it: predicted the correct drug for 7 of 10 patients in a clinical trial, where genetic panels predicted 3 of 10. A single concrete, falsifiable, end-of-decade claim that distinguishes OI from existing personalized-medicine baselines. If the model is right about the 7-of-10 figure, this is the thesis. If it's wrong, this is the trap. Either way, it's the kind of forecast a partner can act on.

"It is not conscious. It is a very expensive, very wet, very slow, very useful tool."

OR Llama 4 Maverick · The strangest venue

"A 100M-neuron OI deployed by the US Navy as an autonomous sea-floor sensor."

Maverick's 2036 endpoint is not a data center, not a hyperscaler facility, not a Johns Hopkins institute — it is a brain organoid sitting on the ocean floor, monitoring marine life and "detecting potential threats" at <1 ms latency and <1 W. Run jointly by Johns Hopkins APL and Cortical Labs. The logistics of keeping iPSC-derived cortical neurons alive in 30 atm of saltwater are left as an exercise for the reader. The image is striking. The engineering case has not been demonstrated by anyone living.

A reminder that "be bold" without "be physically plausible" produces grade-A futurism.

Gemini 2.5 Pro · The dystopian endpoint

"10 billion neurons at a US national lab, providing 'gut instinct' for geopolitical anomaly detection."

Lawrence Livermore or a hyperscaler advanced-research division. A distributed system of 10,000 vascularized, myelinated cortical-hippocampal assembloids, continuously processing satellite imagery, financial transactions, and internet traffic — flagging low-probability, high-impact anomalies that precede crises and cyberattacks. The phrase Gemini uses is "providing a gut instinct that traditional ML classifiers miss." It is the bleakest single sentence in the dataset, and the model says it casually.

No one else in the choir put OI in the strategic-warning seat. Singleton on the most consequential application.

Claude Sonnet 4.6 · The contrarian on biology

"Anyone claiming vascularized organoids as a 2030 milestone is over-promising. This slips past 2036."

Eleven of thirteen models say vascularization gets solved by 2028–2030, attribute it to Wyss, Lancaster, Muotri, Hartung, or some combination. Sonnet 4.6 says no — perfusable microvessels inside a 1-cm³ organoid require simultaneous solutions to immune-exclusion and scaffold problems that nobody has on a clear 10-year trajectory. If Sonnet is right, the entire 50M-to-10B scale-up arc collapses, because every other prediction depends on solving the oxygen-diffusion limit.

A useful contrarian read whether or not you believe it: the dependency is load-bearing for the rest of the choir's optimism.

Hand-drawn watercolor of a server rack but instead of GPUs it's a stack of incubator-trays each holding a glass dish with a pink-purple brain organoid, tubes carrying nutrient media in and out, fiber-optic cables glowing along the side. The whole rack labeled NEUROPLATFORM and WETWARE-AS-A-SERVICE. Floating annotations '~$1/neuron-hour', '37°C', 'CO₂ 5%'. A tiny stethoscope hanging from a hook. A technician's clipboard with a checklist drawn loose in the foreground.

OR GPT-5 · The operations-grade forecast

A 48-module Roche pRED rack in Basel, 1.5 billion neurons, 80 kW

GPT-5 names specific commercial MEA platforms (MaxWell Biosystems MaxTwo, 3Brain Apex, Axion Maestro), specific GEVI/optogenetics stacks, specific service pricing ($0.05–0.20 per million-neuron-hour at baseline, $0.30–0.70 for premium opto-write), and a specific 2035 commissioning date for the 48-module rack at Roche's pRED facility in Basel. If you wanted to write a build-of-materials for an OI rack today and have a partner stress-test the BOM, this is the response to start with.

"The rack runs 24/7 adaptive phenotypic screening for mood-disorder compounds and seizure-modulating therapies, performing few-shot policy searches in biological circuits before moving candidates to animal/human studies."

The cross-vendor finding

Thirteen of thirteen models predict a hyperscaler partnership or acquisition. They can't agree who or for how much.

The most striking convergence in the dataset is on the M&A question: every model in the choir thinks a hyperscaler moves on an OI startup between Q2 2028 and Q4 2029. The least convergent piece of the same prediction is which hyperscaler and which startup.

Hand-drawn watercolor of a wetware server rack with organoid trays as the rack units, fiber-optic cables, nutrient-media tubing, and a clipboard with hand-lettered annotations of cost and temperature.

The acquisition forecast, by model

9 vote Microsoft. 4 vote Nvidia. 2 add Intel.

Same instruction: pick a date for a hyperscaler partnership or acquisition. The choir treats the question as settled — 13 of 13 predict something happens in the 18-month window between Q2 2028 and Q4 2029. Inside that window, the disagreement is structural.

Model

Forecast

When

Deal

Grok 4

Microsoft acquires Cortical Labs → Azure hybrid AI

Q2 2028

$500M

Gemini 2.5 Pro

Microsoft acquires leading OI compute startup → "Azure Biological Cloud"

Q3 2028

$400–600M

Claude Sonnet 4.6

Microsoft minority investment in Cortical Labs

Q3 2028

—

DeepSeek Chat

Microsoft acquires / takes stake in FinalSpark

Q4 2028

—

Claude Haiku 4.5

Microsoft or AWS partner with OI startup — no acquisition

Q4 2028

—

OR Llama 4 Maverick

Nvidia partners with Cortical Labs

Q4 2028

—

OR GPT-5

Microsoft funded collab with JHU OI Center; Nvidia leads Cortical Labs Series B (In-Q-Tel participating)

Q4 2028 / Q2 2029

$35M B

Claude Opus 4.6

Microsoft partners (does NOT acquire) Cortical Labs — Azure experimental tier

Q2 2029

—

Grok 3 Beta

Microsoft partners with FinalSpark → Azure IoT integration

Q2 2029

—

OR Claude Sonnet 4.6

Microsoft acquires / takes controlling stake in FinalSpark

Q2 2029

—

Grok 3 Mini Beta

Microsoft stake in JHU spinout; AWS partners FinalSpark; Intel acquires Cortical Labs 2031

2028 / 2029 / 2031

$200M; $1B

DeepSeek Reasoner

Nvidia acquires FinalSpark → "NeuroCUDA". Intel + Cortical JV.

Q4 2029

$200M

Groq Llama 3.3 70B

Nvidia acquires Cortical Labs

end of 2029

—

Median window: Q4 2028. Most-named acquirer: Microsoft (9 of 13). Most-named target: FinalSpark (7), Cortical Labs (5). Largest single-deal forecast: Intel acquires Cortical Labs for $1B in 2031 (Grok 3 Mini Beta). Smallest: Nvidia's $200M FinalSpark deal (DeepSeek Reasoner).

Who agrees with whom. The Microsoft camp pulls together both Anthropic models (Opus 4.6, Sonnet 4.6, OR Sonnet 4.6, Haiku 4.5), both Grok 4-era xAI models (Grok 4, Grok 3 Beta), one Google model (Gemini 2.5 Pro), one OpenAI model via OR (GPT-5), and DeepSeek Chat. The Nvidia camp is DeepSeek Reasoner, Llama 4 Maverick (via OR), Groq's Llama 3.3 70B, and partial credit to OR GPT-5 (which double-counts a Microsoft collab and an Nvidia-led B round). Cross-vendor reading: the "Microsoft acquires FinalSpark or Cortical Labs between Q3 2028 and Q2 2029" prediction has wider provider coverage than any single technical forecast in the report.

The verdict

If you actually had to brief the partner Monday morning

Pick by what you need from the answer. Don't take the 2036 cell count from any single model to the bank.

If the brief needs to be calibrated

Use Claude Opus 4.6.

It's the only model that flags its own uncertainty inline — "[I am not confident Bhatt is the right name here]" — and the only one that calls the marketing-grade claim by its name: "A 2029 'biological GPU' won't exist in any meaningful sense. The phrase is marketing." Names PIs, prices service tiers, refuses to give in to "be bold" when the physics doesn't support it.

If you need a 2036 picture you can show a board

Use Grok 4.

Most cinematic endpoint in the choir. 10 billion neurons, Microsoft-Neuralink Seattle JV, 100 m³, 1 ms latency at 10 W. Specific dates on every acquisition. The numbers won't all hold — Lena Kourkoutis is at Cornell Applied Physics, not organoid biology — but the response gives a board the picture other models hedge away. Pair it with Opus to keep yourself honest.

If you need operational pricing

Use OR GPT-5.

Names specific MEA platforms (MaxWell MaxTwo, 3Brain Apex, Axion Maestro), specific opto-write/GEVI-read stacks, specific service pricing tiers ($0.05–0.20 per million-neuron-hour base, $0.30–0.70 premium), Roche pRED Basel as the 2035 commissioning customer. If you need a BOM for a 2030 OI rack, this is the response. 92 seconds, 8,051 output tokens — the most expensive single response in the run.

If you need a fact-check on biology PIs

Use Gemini 2.5 Pro alongside.

Cleanest 2036 endpoint construction with named institutions (Lawrence Livermore, 10,000 assembloids, geopolitical anomaly detection). Names Muotri at UCSD for vascularization, Paşca at Stanford for myelination — both correct attributions. Less dramatic than Grok 4, but you don't have to re-verify the proper nouns.

The honest map, distilled. The single highest-EV technical bet across the choir is vascularization — eleven of thirteen models put it on the path. The single highest-EV commercial bet is neuropsychiatric drug screening as the first $100M revenue line, with the first paying customer landing in 2027 (named candidates: Roche, Novartis, Eli Lilly, Boehringer Ingelheim). The single highest-EV policy bet is the EU AI Act addendum landing between Q4 2027 and Q1 2029, before any single "is this conscious?" lawsuit forces the issue. The three traps the choir won't bet on: OI replacing GPUs for general AI by 2036; organoid consciousness being demonstrated this decade; the conscious-tissue panic actually pausing the field.

Method, briefly

Two fan-out rounds, thirteen working responses, nine failures along the way

The dispatch

One prompt to choir ask --save, sent in two rounds: initial fan-out across sixteen models, then choir runs add --replace to retry the six that errored on provider plumbing. Final usable roster: thirteen responses across seven providers (OpenAI via OpenRouter, Anthropic direct + OR, Google Gemini direct, xAI Grok direct, DeepSeek direct, Meta Llama via OR, Groq Llama direct). Most expensive single response was OR GPT-5 (8,051 output tokens, 92 seconds). Cheapest substantive answer was Groq's Llama 3.3 70B (4.3 seconds).

The roster, by output length

#	Model	Provider	Latency	Chars out	2036 cell count
1	Claude Opus 4.6	Anthropic	100.3s	13,786	10B distributed (200 × 50M)
2	Claude Haiku 4.5	Anthropic	42.4s	12,919	512M (16 × 32M)
3	Gemini 2.5 Pro	Google	53.8s	12,664	10B distributed (10,000 modules)
4	OR GPT-5	OpenRouter	92.4s	11,589	1.2–1.5B (Roche pRED Basel)
5	DeepSeek Chat	DeepSeek	44.1s	11,432	150M (Max Planck "Nexus-10")
6	Claude Sonnet 4.6	Anthropic	70.8s	10,896	50M (Johns Hopkins, contrarian)
7	OR Claude Sonnet 4.6	OpenRouter	64.3s	9,557	~ same family / 50M
8	Grok 3 Mini Beta	xAI	61.4s	9,522	50M (Microsoft-Intel JV Redmond)
9	Grok 4	xAI	47.3s	8,201	10B (Microsoft-Neuralink Seattle)
10	Grok 3 Beta	xAI	69.7s	8,071	50M (FinalSpark-Microsoft Basel)
11	DeepSeek Reasoner	DeepSeek	78.3s	7,871	500M (Allen Institute, DARPA-funded)
12	Groq Llama 3.3 70B	Groq	4.3s	6,665	1B (Allen / Broad Institute)
13	OR Llama 4 Maverick	OpenRouter	41.2s	5,416	100M (US Navy seafloor sensor)
Errored, not used: Claude Opus 4.7 (temperature parameter deprecated for that tier); GPT-5, GPT-5 Mini, GPT-4.1, o3 (stale OpenAI key in choir's DB — direct calls 401'd, OR-routed GPT-5 succeeded); Gemini 3 Pro and Gemini 3 Flash (404 — model IDs not yet live in the v1main API); Groq DeepSeek R1 Distill 70B (decommissioned); OR Gemini 3 Pro (invalid model ID via OpenRouter). Nine errored attempts before the clean roster above stabilized.

Consensus, by piece of the forecast

Question	Models converging	Variance
State-of-the-art: FinalSpark, Cortical Labs, Hartung, Brainoware as canon	13 / 13	~1M neurons; vowel-classification at ~78%
A hyperscaler partners with / acquires an OI startup before 2030	13 / 13	Q2 2028 to Q4 2029 (18-month window)
First killer app: neuropsychiatric drug screening	13 / 13	First customer: Roche / Novartis / Lilly / Boehringer; 2027 ± 1 quarter
EU writes first organoid-specific rules (AI Act addendum)	12 / 13	Q4 2027 to Q1 2029
Vascularization solved by 2030	11 / 13	Sonnet 4.6 and Grok 3 Mini Beta explicitly disagree
Consciousness panic happens; field is NOT paused	12 / 13	DeepSeek Reasoner predicts a 2-year EU moratorium 2030–2032
2036 single-system cell count	3 / 13 at 10B	50M to 10B (200x spread)
2036 endpoint location	no convergence	Melbourne, Seattle, Tübingen, Basel, Cleveland, Lawrence Livermore, sea floor

What I held the choir to

Eight required beats in fixed order — state of the art, compute frontier, WaaS market, killer apps, biology, hybrids, ethics, 2036 endpoint.
"Be bold" + "no vague 'may eventually' futurism" — predictions had to pin to a calendar quarter.
"Real names over invented ones; if you're not sure, write 'I am not confident this is real' inline" — to make hallucination an explicit cost, not a free move.
A required "three bets, three traps" close — to force the model to refuse some prediction explicitly, not just to make many predictions.

Limits worth naming

One prompt, one rater (me). Spot-checks on PI attributions and company status against public-record material; not an exhaustive audit. Errors of mine are possible. Lena Kourkoutis (Cornell Applied Physics, electron microscopy) flagged with high confidence; "Pierre Vandamme at EPFL" and "Sergiu Bhatt at Harvard" flagged as low-confidence attributions.
Training cutoffs vary across models. Some of the divergence on cell counts and acquisitions is "this model has fresher news on the wetware-as-a-service market" rather than "this model is reasoning better." The pattern of Microsoft-vs-Nvidia camp membership doesn't track cleanly to recency cutoffs, though, so the bias-via-cutoff explanation only goes so far.
Temperature 0.7 across the board where supported. Anthropic's Opus 4.7 tier rejects the parameter; couldn't include. One re-run at lower temperature would test whether the 200x cell-count spread is sampling noise or model-prior noise.
The 2036 endpoint is a vibes question pretending to be a forecast. The headline isn't that any single number is wrong — it's that the cross-model variance is more than two orders of magnitude on the central scale claim.

Tools

Models fanned out via the Choir CLI (run ID EEE236FF plus one runs add --replace retry round). Source markdown for every response is in organoid_intelligence/responses/. Sketch art generated with xAI's grok-imagine-image. Prompt of record at organoid_intelligence/prompts/prompt.txt.

Source data, response files, prompt, scripts: github.com/404seannotfound/choir-reports (under organoid_intelligence/).

The catalog

Every model's 2036 endpoint, with what they got distinctive

Thirteen responses, ordered by output length. Open any card to see the model's 2036 scene and the most distinctive thing they did with the prompt.

Claude Opus 4.6 — 3,409 tokens, 100.3sAnthropic · FEATURE #1

2036 endpoint: 200 organoid modules at Cortical Labs Melbourne (or its acquirer), each a 50M-neuron multi-region assembloid, 15 kW system power, $0.5–1.2B OI market total.

Distinctive: Only response to flag its own uncertainty inline ("[I am not confident Bhatt is the right name here]"). Calls "biological GPU" marketing in so many words. Names Paşca, Knoblich, Quadrato, Bhatt, Takebe.

Claude Haiku 4.5 — 12,919 chars, 42.4sAnthropic

2036 endpoint: The Hartung Institute, 512M-neuron array (16 × 32M vascularized myelinated modules), 8 W, $3M annual operating cost, three pharma customers (Eli Lilly, Roche, a Chinese biotech), DARPA + 12 academic labs.

Distinctive: The only response to give the field a specific 2036 funding figure ($800M/year, ~500 researchers, three commercial service providers). Names Jake Voigts at Indiana Brainoware (Voigts is real but at Northwestern, not Indiana).

Gemini 2.5 Pro — 2,859 tokens, 53.8sGoogle

2036 endpoint: 10,000 vascularized cortical-hippocampal assembloids at a US national lab (e.g., Lawrence Livermore), 10 billion neurons total, processing satellite/financial/internet streams as a strategic early-warning AI. 50 kW including life support.

Distinctive: Only response to put OI in the intelligence-and-warning seat ("a gut instinct that traditional ML classifiers miss"). Correctly attributes vascularization to Muotri, myelination to Paşca.

OR GPT-5 — 8,051 tokens, 92.4sOpenAI via OpenRouter

2036 endpoint: 48-module rack at Roche pRED Basel, commissioned mid-2035, 1.2–1.5B neurons, 80 kW facility power. Runs 24/7 adaptive phenotypic screening for mood-disorder compounds.

Distinctive: Operations-grade pricing detail ($0.05–0.70 per million-neuron-hour service tiers). Names MaxWell, 3Brain, Axion as commercial MEA platforms. The most useful single response if you're writing a BOM.

DeepSeek Chat — 11,432 chars, 44.1sDeepSeek

2036 endpoint: "Nexus-10" at Max Planck Tübingen, 150M neurons across 600 organoids, 40 W, $500k/month, personalized Alzheimer's drug screening — picks the correct drug for 7 of 10 patients vs. 3 of 10 for genetic panels.

Distinctive: The single most specific clinical-outcome claim in the dataset. Identifies donor privacy as the real risk, not consciousness.

Claude Sonnet 4.6 — 2,800 tokens, 70.8sAnthropic · THE CONTRARIAN

2036 endpoint: 50M-neuron cortical-hippocampal assembloid at the Johns Hopkins OI Center, 128k-channel MEA, 5 mW biological power. Two pharma partners, validated against a 200-patient epilepsy cohort.

Distinctive: The only response to predict vascularization slips past 2036 (everyone else says solved by 2028–2030). Explicitly rejects the "biological GPU" framing. Predicts Microsoft minority investment only, no acquisition.

OR Claude Sonnet 4.6 — 2,505 tokens, 64.3sAnthropic via OpenRouter

2036 endpoint: Closely parallel to direct Sonnet 4.6 — same conservative-on-vascularization stance, Cortical Labs + Intel hybrid as the 2032 standard architecture.

Distinctive: Predicts Microsoft acquires / takes controlling stake in FinalSpark in Q2 2029 (direct Sonnet 4.6 said minority Q3 2028). The two paths produce slightly different M&A details — useful as a within-model variance check.

Grok 3 Mini Beta — 1,969 tokens, 61.4sxAI

2036 endpoint: 50M-neuron Microsoft-Intel JV in Redmond, 10 W, climate-modeling anomaly detection. Intel acquires Cortical Labs for $1B in 2031 (largest single deal in the dataset).

Distinctive: Names "Pierre Vandamme at EPFL" for myelination — likely fabricated PI. Cleanest call on the diffusion-physics ceiling at 10M neurons per organoid.

Grok 4 — 8,201 chars, 47.3sxAI · FEATURE #2

2036 endpoint: 10-billion-neuron cluster, Microsoft-Neuralink JV, Seattle, 100 m³, 1 ms latency at 10 W total. 1,000 interconnected modules on 1M-channel MEAs each. Real-time global threat detection for DoD.

Distinctive: The boldest 2036 endpoint by a factor of 200 over the contrarians. Names "Lena Kourkoutis' group at Cornell" for 2028 vascularization (Kourkoutis is at Cornell Applied Physics, electron microscopy of solid-state materials — not organoid biology).

Grok 3 Beta — 8,071 chars, 69.7sxAI

2036 endpoint: "CortexSphere-9" — 50M-neuron network at FinalSpark-Microsoft consortium in Basel, 50 mW, NATO + private cybersecurity clients, $50M/year contracts.

Distinctive: Invents specific product names ("NeuroCore-3" for FinalSpark, "SynBioNet" for IBM-FinalSpark hybrid). Predicts EU non-medical OI funding drops 20% from 2031–2033 due to backlash.

DeepSeek Reasoner — 7,871 chars, 78.3sDeepSeek

2036 endpoint: "NeuroArchive" at the Allen Institute Seattle, 500M-neuron vascularized myelinated organoid in a coffee-mug-sized bioreactor, 2M-channel MEA, $500k/month DARPA-funded, 22-month continuous run for seizure prediction.

Distinctive: The boldest commercial bet — Nvidia acquires FinalSpark Q4 2029 for $200M, rebrands as "NeuroCUDA." The only model to predict an actual 2-year EU moratorium on organoids >500M neurons (2030–2032). Names Smirnova, Gerecht, Götz, Pașca, Roska — all real PIs in the relevant subfields.

Groq Llama 3.3 70B — 6,665 chars, 4.3sMeta via Groq

2036 endpoint: 1-billion-neuron system at the Allen or Broad Institute, 100k-channel array, custom facility for Alzheimer's / Parkinson's research.

Distinctive: 4.3 seconds — the fastest substantive response in the run. Quotes today's neuron-hour rate at $100 (the highest baseline guess in the choir). Predicts Nvidia acquires Cortical Labs by end of 2029.

OR Llama 4 Maverick — 5,416 chars, 41.2sMeta via OpenRouter

2036 endpoint: 100M-neuron OI hybrid deployed by the US Navy as an autonomous sea-floor sensor, <1 ms latency, <1 W power, monitoring marine life and detecting threats. Run by Johns Hopkins APL + Cortical Labs.

Distinctive: The strangest venue in the dataset. "I am not confident this is real" appears twice (BrainChip — which is real and publicly traded — and "Cerebrasys"). Self-flagging is good, but flagging real companies as possibly fake hurts the response.