AMD's $8.2 billion World Labs deal has a share count that does not exist yet

FACT (CONFIRMED, AMD primary). AMD issued a press release at 16:05 EDT on Monday 2026-09-28, distributed via GlobeNewswire, stating it "has entered into a definitive agreement to acquire World Labs, an AI model and research lab led by AI pioneer Dr. Fei-Fei Li."

FACT (CONFIRMED, SEC primary — the detail the press release omits). AMD's Form 8-K under Item 3.02 (Unregistered Sales of Equity Securities), accession 0000002488-26-000182, period of report 2026-09-26, filed 2026-09-28 at 16:16:47 ET, states verbatim:

"On September 26, 2026, Advanced Micro Devices, Inc. (the 'Company') entered into an Agreement and Plan of Merger (the 'Merger Agreement') to acquire all of the equity interests in World Labs Technologies, Inc. ('World Labs') (the 'Acquisition') for a total purchase price of approximately $8.2 billion to be paid in shares of the Company's common stock, par value $0.01 per share ('Common Stock'), to stockholders of World Labs, subject to customary adjustments."

"At this time, the number of shares to be issued in connection with the Acquisition is not known and will be calculated based on the daily volume-weighted average price of a share of Common Stock on the Nasdaq Global Select Market over the ten (10) consecutive trading-day period ending on and including the second trading day immediately preceding the closing date of the Acquisition."

"The Company intends to issue the shares of Common Stock in reliance upon the exemptions from registration afforded by Section 4(a)(2) of the Securities Act of 1933 … and/or Rule 506 of Regulation D … The issuance is not being conducted in connection with a public offering."

The filing is signed by Ava Hahn, Senior Vice President, General Counsel and Corporate Secretary. It was accepted 11 minutes and 47 seconds after the press release timestamp.

FACT (CONFIRMED, AMD primary). Deal structure and leadership, per the press release: the all-stock transaction is valued at approximately $8.2 billion, expected to close by the end of 2026, subject to regulatory approvals and other customary closing conditions. Following the close, the World Labs team "will continue to focus on advancing AI model research," and Dr. Fei-Fei Li "will join AMD as executive vice president and chief scientist, reporting to Dr. Lisa Su."

FACT (CONFIRMED, AMD primary). What World Labs is, in AMD's own description: "Headquartered in San Francisco, World Labs develops spatial-intelligence models that generate, reconstruct and simulate interactive 3D environments from text, image and video inputs, as well as technology for robotic learning and simulation."

FACT (CONFIRMED, quoted principals).

"Building the compute platforms for the next generation of AI requires a deep understanding of how models are evolving," said Dr. Lisa Su, chair and CEO, AMD. "Fei-Fei and the World Labs team bring exceptional research leadership and model expertise. Together, we can use that insight to develop the hardware, software and systems that will power the next generation of AI and strengthen the open AI ecosystem."

"Advancing the next generation of AI technology requires close collaboration across model research, systems and compute," said Dr. Fei-Fei Li, co-founder and CEO, World Labs. "Joining AMD will give our team the resources and engineering depth to accelerate our research and help define the infrastructure needed for the next era of AI."

FACT (CONFIRMED, both companies). World Labs' own announcement blog adds a personnel fact not in AMD's release: Justin Johnson and Ben Mildenhall will work with Li to continue leading the World Labs team, with the stated goal of building "an end-to-end open AI ecosystem spanning hardware, software, platforms, and widely accessible open models."

FACT (CONFIRMED, Li, personal essay "To Seek a Newer World", 2026-09-28). World Labs was founded with Mildenhall and Johnson "at the beginning of 2024." Li states:

"To do this requires scaling our efforts, widening our reach, and getting closer to the hardware. Without having a focused hardware effort, AI is hobbled in efficiency. And scale."

and that AMD CEO Lisa Su "was an early investor and believer in our mission." Li commits the combined organisation "to providing the best open models and platforms for running AI end to end from hardware to software to models to the data platforms needed to fuel it all."

FACT (CONFIRMED, INDEPENDENTLY VERIFIED across five outlets). Reuters, CNBC, TechCrunch, Forbes and Data Center Knowledge all report $8.2 billion, all-stock, expected close by end of 2026, and Li as EVP and chief scientist reporting to Su. No outlet retrieved reported a different value, structure, or reporting line.

FACT (CONFIRMED, market reaction). AMD closed down 3.6% at $607.87 on Monday 2026-09-28, and was up 0.1% in after-hours trading following the announcement (Forbes, 2026-09-28 17:05 ET). AMD shares were up 183% year-to-date at that point, having traded around $223 at the start of 2026. Investor's Business Daily reported on 2026-09-29 that AMD "earned positive marks from Wall Street analysts" for the deal while the stock "dipped on Tuesday."

INDEPENDENT EVIDENCE — the pre-existing relationship, which materially shapes the story. This was not an arms-length acquisition of a stranger.

• AMD was an investor in World Labs. Reuters: "World Labs raised $1 billion in a funding round earlier this year that included an investment from AMD" (round closed 2026-02-18; other named investors Nvidia, Autodesk, Autodesk investing $200M and taking an adviser role, Emerson Collective, Fidelity, Sea). Li separately describes Su as "an early investor."
• AMD and World Labs had a technical partnership beginning in 2025, "starting with model training and inference optimization on AMD GPUs" (Li, World Labs blog and her essay; also TechCrunch and Data Center Knowledge).
• Li appeared on stage at AMD's CES 2026 keynote on 2026-01-05, demonstrating World Labs' Marble model with Su (CNBC; Reuters photo caption of the same event).

The AI chip roadmap and the frontier-model roadmap are now the same organisation's budget.

Before this event, the structural convention across the whole industry was that model labs rented silicon. AMD sold accelerators to whoever was building the frontier; it did not know what those customers' models would need next year, and had no contractual reason to care beyond a sale. NVIDIA's position was different — it had spent years building a full stack (per Data Center Knowledge: Cosmos, Isaac, Omniverse, Project GR00T) precisely because shipping chips and understanding the workload are the same discipline.

On 2026-09-28 AMD bought the workload. FACT (CONFIRMED): the press release's stated rationale is not revenue. It is foresight — "World Labs' expertise in developing advanced models will give AMD deeper insight into how workloads are evolving and help shape its future technology roadmaps," framed around "reasoning, robotics, simulation and physical AI."

INTERPRETATION — the inversion is structural, not cosmetic. The discovery framing ("a GPU vendor buying a frontier AI lab to steer silicon") is directionally right but understates the mechanism. The specific inversion is this: AMD will now be able to design an instruction set, memory subsystem or interconnect to a workload it co-owns, years before that workload shows up as a purchase order. Matt Kimball (VP and principal analyst, Moor Insights & Strategy) put the mechanism precisely:

"This isn't about AMD buying a model company. This is about being able to understand the requirements of compute and memory – and the relationship between the two as the next generation of AI emerges. … I like the move. To me, it is AMD looking a few years out and capturing those needs today – needs that get baked into low-level architectural decisions and designs."

Three second-order changes follow:

1. The "second-largest acquisition" label is the wrong headline; "AMD's most consequential lab-to-silicon commitment since Xilinx" is the right one — but the differentiating fact is the reporting line, not the size. Li reports to the CEO, not to a business-unit president. An EVP/chief scientist with a direct line to the CEO and a mandate to shape hardware is a different instrument than a business-line acquisition. INTERPRETATION.
2. AMD bought a roadmap it had already been financing. As a shareholder in the February 2026 round and a 2025 technical partner, AMD had privileged access to World Labs' compute plans for roughly a year before signing. INTERPRETATION: this is not a governance scandal — both sides disclosed it — but it does mean the "$8.2B for a company" framing understates what a dollar of that price bought. Part of it bought information AMD was already entitled to. This is an interpretation about deal economics, not an allegation.
3. Physical AI got a second silicon vendor with a first-party model story. Until now, an enterprise evaluating world models for robotics or simulation had essentially one silicon-integrated option (NVIDIA) and a set of independent model labs that rented chips. That set is now materially smaller.

To understand why this deal, you need to understand why a world model's compute profile is different enough to be worth buying a lab for.

5.1 What a world model is, in the target's own architecture

FACT (CONFIRMED, World Labs primary — Atlas post, 2026-09-01). Atlas is described as a multimodal autoregressive diffusion transformer, "pretrained from scratch to natively operate on text, images, video, and 3D." The four architectural properties and why each exists:

• Multimodal — natively processes text, images, camera poses, and 3D depth maps; videos are represented as sequences of images. Every image and depth map is conditioned on an explicit camera pose.
• Autoregressive — operates on sequences, generating one element at a time conditioned on earlier ones. "Each task is just a different kind of sequence, where inputs are followed by outputs."
• Diffusion — a rectified flow model that generates by gradually denoising. "can naturally trade off speed and quality by varying the number of denoising steps used during inference."
• Transformer — "consists primarily of large matrix multiply operations and is well-adapted to modern hardware. It is a robust backbone for world modeling."

The differentiator is what World Labs calls spatial context: "each image is grounded at a 3D position in space." Unlike an LLM, where context is a token sequence, Atlas's context is geometrically indexed. Data Center Knowledge's Kimball made the same point in plain language: "While LLMs focus on the relationship between language, tokens, and information, world models are about environments — geometry, physics, objects, motions, and dynamics."

INTERPRETATION — why this matters for silicon. Three properties of that architecture map onto three different hardware bottlenecks, which is why a chip vendor cannot learn these requirements from order data:

1. Camera pose as a native input type. Atlas does not take a text prompt describing camera motion; it takes precise geometry, "going beyond coarse text-based instructions." INTERPRETATION: that implies a native tensor/attention path for 6-DoF extrinsics and intrinsics, and precision requirements on pose encoding that a general LLM path never exercises. Co-design opportunity.
2. Adjustable denoising steps. World Labs explicitly frames denoising-step count as a speed/quality dial. INTERPRETATION: diffusion inference is a throughput- and latency-sensitive regime with a different arithmetic-intensity profile and a different batching behaviour than autoregressive decode. The same GPU that is memory-bound on token decode can become compute-bound on diffusion rollout. This is exactly the "relationship between compute and memory" Kimball named.
3. Explicit 3D outputs. Atlas "natively operates on both 2D image frames and 3D depth maps, enabling it to output worlds as point clouds or 3D Gaussian splats," and the same 3D Gaussian splat representation is used in Marble. INTERPRETATION: this is a rasterisation-and-rendering workload sitting downstream of a generation workload — a different pipeline shape entirely, and one that World Labs already exposes as an open-source renderer (Spark, for THREE.js).

5.2 The simulation loop, and why it is the real prize

FACT (CONFIRMED, World Labs primary — Real-to-Sim-to-Real post, 2026-07-28). World Labs acquired SceniX on 2026-07-21, a robotics and simulation company building a real-to-sim-to-real (R2S2R) engine. The documented results:

• Policies trained entirely in simulation with zero real-world training data transferred directly to real robot platforms — ALOHA (bimanual box packing, cube handover), RB-Y1 (power-cord manipulation around a refrigerator), YAM (cable manipulation), Flexiv (test-tube transfer), xArm (marker and pencil singulation from clutter).
• Multiple demonstrations show robots "operated autonomously for one hour without intervention."
• The evaluation claim is the commercially important one: each checkpoint evaluated on 2,000 simulated trials (1,000 in-distribution + 1,000 out-of-distribution) versus 100 real-world trials (50 ID + 50 OOD), with simulated evaluation reported to preserve real-world policy rankings and to reveal similar success/failure regions.

World Labs' own framing of why this is the bottleneck: "Unlike internet data used to train language models, robot experience is expensive to collect and difficult to control," and "Robot development iterates orders of magnitude more slowly than language-model development because policy evaluation remains heavily tied to physical hardware."

INTERPRETATION. The compute implication is the point of the acquisition. R2S2R turns robotics from a data-collection cost into a simulation-throughput cost — millions of environment steps, all needing generation and rollout. That is a sustained, large, forecastable compute demand of a kind that is visible in advance, which is precisely the kind of demand AMD's cautionary statement flags as material ("loss of a significant customer"; "memory conditions"). The buyer's rationale in the press release — "help shape its future technology roadmaps" — makes most sense against this workload, not against creative 3D generation.

5.3 The pricing mechanism, which is the part everyone is glossing over

FACT (CONFIRMED, SEC primary). Because consideration is priced off a trailing 10-trading-day VWAP ending two trading days before closing, the exchange ratio is a function of AMD's own share price at a future date, not today.

INTERPRETATION — the asymmetry this creates. World Labs' shareholders are implicitly short AMD stock from closing until settlement: the lower AMD trades, the more AMD shares they receive for the same ~$8.2B headline value, and the more existing AMD shareholders are diluted. AMD's stock fell 3.6% on announcement. At the announcement-day close of $607.87, ~$8.2B implies roughly 13.5 million shares. A 10% drawdown before the pricing window implies ~15.0 million shares (+11.1%); a 25% drawdown implies ~18.0 million shares (+33.3%). These are my arithmetic illustrations on AMD's closing price, not forecasts, and the actual calculation uses VWAP over a trailing ten-day window, not a single close. The mechanism's existence is CONFIRMED; the share count is not knowable until closing.

INTERPRETATION. This has a second-order consequence that no retrieved source notes: the trailing-window design means World Labs' shareholders bear AMD's share-price risk for the window immediately before closing, which is precisely the period during which leak-driven or market-driven volatility in AMD would be most likely. Whether that is protective or adversarial depends on the collar and cash-out provisions in the merger agreement — which have not been made public. This is listed in §14 as an open question, not asserted as a finding.

INTERPRETATION — the durability test for a chip vendor is whether it can anticipate workloads, and anticipation is what acquisitions buy. NVIDIA's defensibility in physical AI rests on owning the stack: silicon plus Cosmos, Isaac, Omniverse and GR00T. AMD's answer was, until this week, to rent demand. This deal buys the ability to specify demand. Kimball's framing is the correct one to carry forward: "AMD looking a few years out and capturing those needs today — needs that get baked into low-level architectural decisions and designs."

INTERPRETATION — the price signals what AMD thinks physical AI is worth, and it is large enough to matter. $8.2B is roughly 0.71× AMD's entire Q2 2026 revenue ($11.5B) and about 3.6× Q2 net income ($2.3B) (AMD Q2 2026 results, 2026-08-04; my arithmetic). Against a company whose data centre revenue more than doubled year-on-year to $6.7B — 58.3% of total revenue, per AMD's own Q2 release and my calculation — this is a serious commitment, not a rounding error. And it is all-stock, so it consumes no cash; it consumes dilution.

INTERPRETATION — the premium is what makes it an acquisition rather than a consolidation. World Labs' last reported private valuation was approximately $5 billion (Bloomberg, reported January 2026, published by Reuters 2026-02-18). $8.2B against $5B is 1.64× — a ~64% premium (my arithmetic; the valuation figure is reported, not company-confirmed, and the 8-K discloses no reference price). A 64% premium roughly seven months after a round at that mark is a strong signal that AMD is not buying a balance sheet.

INTERPRETATION — talent is an explicit line item in the price. Burke (Futurum Group): "You can't hire a more accomplished researcher than Dr. Fei-Fei Li, and access to talent is the biggest advantage in AI training." Kimball, blunter: "Fei-Fei is a rock star. You could argue that she is as much of the target as the company is." Both are analyst opinion, not company disclosure, and AMD has not said this. But the reporting line is the evidence: Li reports to the CEO directly, not through a segment P&L.

INTERPRETATION — for the ecosystem, the model-lab-rents-chips relationship just became optional. Independent world-model labs were the mechanism by which a robotics or simulation startup got frontier spatial intelligence without betting on one hardware vendor. World Labs was arguably the most credible such lab. That option is now materially narrower.

INTERPRETATION — "world models" graduated from research curiosity to infrastructure category. Brendan Burke cites AMD's own estimate that physical AI "could reach $200 billion by 2035" (COMPANY CLAIM, relayed by an analyst, not independently verified here). Even discounting the number, the direction of travel — a chip vendor building its roadmap around a model class — is the real signal. World Labs' own taxonomy of world models spans "renderers, simulators, and planners"; AMD has now bought one of each capability at once.

INTERPRETATION — newly possible because of this event:

1. Silicon co-design against a first-party spatial workload. Camera-pose conditioning, rectified-flow diffusion with an adjustable step budget, and explicit 3D/point-cloud/splat outputs are now requirements AMD can design to, with a research team available to validate. This is the core of the deal.
2. A credible second source for physical-AI tooling. NVIDIA had Cosmos (open-weight), Isaac, Omniverse and GR00T. AMD can now offer an in-house alternative rather than pointing buyers at a partner. For an enterprise running both NVIDIA and AMD fleets, that changes the second-vendor conversation.
3. A bridge from simulation to silicon for robotics. R2S2R's claim is that simulation becomes a high-throughput evaluation layer for robot policies. If that holds, simulation-throughput becomes a major recurring compute line — and it is a line AMD can now see coming rather than infer from a hyperscaler's capex plan.
4. A named, citable precedent for hardware-vendor lab acquisition. Where the prior pattern was chip companies buying chips companies (Groq, per CNBC's account of NVIDIA's $33B combined spend), AMD bought a model lab. Expect this to be imitated or countered.
5. A succession signal that is at least documented. Li becomes EVP and chief scientist reporting to the chair and CEO, joining AMD's senior leadership with an explicit mandate to shape hardware. Characterisation of this as a CEO-succession move appeared in commentary retrieved only at search-snippet level and was not verified against a read article body — see §13. Treat it as interpretation, not reporting.

INTERPRETATION — forecloses:

• A neutral, independent world-model vendor for physical AI. That option is materially reduced. Buyers who wanted World Labs' models without taking a silicon position must now buy from a competitor, or from a company whose roadmap incentives point at AMD GPUs.
• Any presumption that AMD's world-model output will remain independent of AMD's silicon roadmap. The two are now inside one P&L. This is a structural fact, not a criticism — it is also exactly what makes the stack competitive.

1. What is the exchange ratio or reference price, and is there a collar? The 8-K omits all consideration mechanics beyond the VWAP method. Until the proxy or S-4, no one can compute dilution. This is the most answerable and most important question.
2. What is the actual share count, and how has AMD traded into the pricing window? The mechanism is confirmed; the answer is not.
3. Who approves the deal, and has any filing been made? No agency named. If AMD's export-control and national-security exposure makes this a reviewed transaction, that should surface in the proxy.
4. What are the termination fee and the walk-away terms? Determines how credible the end-2026 close really is.
5. Does "chief scientist" carry a hardware-roadmap mandate with budget authority, or is it an advisory title? Li's own essay argues "Without having a focused hardware effort, AI is hobbled in efficiency" — but that argues for AMD having a hardware effort, not for the lab owning one. The mandate's boundaries are undefined.
6. Will Atlas and successors remain open-weight? Both principals committed to open models; NVIDIA already ships open-weight Cosmos. Nothing is contractual. This is the question that determines whether the acquisition beats or joins NVIDIA's stack.
7. What is AMD's physical-AI product and shipping timeline? No roadmap, no date, no model availability commitment. AMD said nothing about what World Labs products become.
8. How many of the World Labs research team are actually retained, and on what terms? Retention packages, non-competes and reporting-line changes are the classic failure point. Nothing is public.
9. Will the acquired lab keep operating as a distinct "frontier research organisation"? Both announcements say yes. Chip-company cadence is the countervailing pressure. Watch for whether World Labs keeps announcing on its own schedule or starts announcing on AMD's.
10. What was AMD's existing stake, and what does it do to the effective purchase price? AMD was an investor in the February 2026 round. Its stake size is undisclosed, so the $8.2B may partly compensate for pre-existing ownership — a real question for AMD shareholders and unanswerable from public sources.
11. Is any independent benchmark of Atlas or R2S2R coming? Every capability claim in the story is vendor-generated. If an independent evaluation of the R2S2R sim-to-real ranking claim appears, it will settle whether the physical-AI thesis has legs.
12. Did the two August–September 2026 AI-M&A deals (NVIDIA/Hugging Face, NVIDIA/Groq) create a competitive countermove? CNBC reported a $33B combined NVIDIA spend. Whether NVIDIA responds to AMD's physical-AI positioning — by shipping something, acquiring, or partnering — is the live strategic question.
13. Did AMD share-price weakness after the announcement reflect deal concerns or unrelated market factors? Forbes attributes the −3.6% to the session in which the deal was announced; IBD reported a further dip on Tuesday. Separating signal from noise requires a longer series than this artifact has.

FACT (CONFIRMED) — the committed dates.

• Closing is expected by the end of 2026, subject to regulatory approvals and other customary closing conditions. No agency named.
• The 10-trading-day VWAP pricing window runs ending on and including the second trading day immediately preceding the closing date — so the share count becomes determinable only at closing.
• Li's EVP/chief scientist role and her reporting line to Su are effective on close, not on announcement.

PREDICTION — the next 90 days, stated as predictions and labelled as such.

• A proxy or S-4 registration statement will be filed, and it will be the first public document containing the exchange-ratio mechanics, any collar, termination fees and the named approval agencies. This is the ordinary course for an all-stock acquisition of this size and it is the single highest-value document still to come. Confidence: high.
• World Labs will continue to publish on its own cadence rather than AMD's, at least initially. Both principals explicitly committed to a distinct "frontier research organization" with continued research focus. Watching whether that holds is the cheapest available read on whether integration is genuine. Confidence: moderate.
• There will be no material AMD physical-AI product announcement inside the window. AMD said nothing about products, and World Labs' own roadmap is Atlas → future Marble versions. Confidence: moderate-high.
• AMD's share price into the pricing window will be a live topic. The mechanism is confirmed and unusual enough that the first analyst note explaining it will generate attention. Confidence: moderate.

FACT (CONFIRMED) — what is already scheduled and observable, unrelated to speculation.

• AMD CEO Lisa Su is delivering the opening keynote at the 2026 OCP Global Summit on 2026-10-12, 4:15 p.m. PDT, San Jose McEnery Convention Center (AMD media alert, 2026-09-29). The alert states Su "will highlight the portfolio of open, full-stack infrastructure for agentic AI that AMD delivers at every scale – from enterprise systems to gigawatt-scale deployments." A silicon vendor keynote is where silicon roadmaps surface, and this one lands 10 days before the research cutoff.
• World Labs is hiring across research and engineering.
• SceniX and the R2S2R programme remain active World Labs workstreams.

INTERPRETATION — the three questions that determine whether this was a good deal.

1. Did the researchers stay? Watch for attrition and for changes to who leads World Labs inside AMD over the next four quarters. The price was substantially a talent price.
2. Did the models stay open? If AMD ships open-weight world models at parity with Cosmos, the acquisition was a strategic masterstroke. If it ships closed, AMD spent $8.2B to build a worse version of NVIDIA's closed stack.
3. Did the silicon change because of it? The testable version: does a future AMD data-centre product show a change traceable to spatial or world-model workloads — camera-pose handling in hardware, a memory subsystem change, a new accelerator block for diffusion? Kimball's claim is that "needs get baked into low-level architectural decisions and designs." By 2028 you will be able to check whether that happened. Until then the deal is a hypothesis.

A chip company bought a model lab, and the most important number in the deal is not the price.

The $8.2 billion is the headline. It is real, it is all-stock, and it is AMD's second-largest acquisition after Xilinx. But three things in the primary record are more informative than the headline, and all three required going past the press release.

First: the deal was signed on Saturday, 2026-09-26, and announced on Monday at 16:05 EDT. AMD's 8-K is dated to the 26th. Small detail, useful discipline — "when did this happen" and "when did we hear about it" are different questions, and only the second one is the answer to "when did this happen."

Second: the share count does not exist yet. Consideration is priced off AMD's own trailing ten-day VWAP immediately before closing. AMD's stock fell 3.6% on the news. That means the number of shares World Labs' shareholders receive is a function of a future price, and it moves inversely with AMD's share price. This is disclosed in a single Item 3.02 filing that carried no headline, and it was not reported by any outlet retrieved. It is also the only part of this transaction with a live, computable, unresolved number — and it is the part an AMD shareholder actually needs.

Third: the deal's rationale is stated in the past tense of silicon design. AMD is not buying revenue. It is buying the right to know, years early, what a new class of workload will need — and that knowledge gets "baked into low-level architectural decisions." That is the honest logic of a second supplier to NVIDIA, and it is a better story than "chip company buys hot AI startup." It is also, unavoidably, a story about a hardware vendor buying an independent voice and making it internal.

What a beginner should carry away, in one sentence: in AI, the line between the companies that make the chips and the companies that make the models has now dissolved on the silicon side too — and that is why the most consequential part of this deal happened in a regulatory filing rather than a press release.

What to be careful about, in one sentence: every impressive number in the story — 94% camera-control preference, zero-real-world-data robot policies, one-hour autonomous operation — was produced by the lab that was sold, and the lab's own blog posts contain the caveats that weaken them.

What remains genuinely open, in one sentence: whether buying a research team gives you a research culture or a hardware roadmap with a research team's name on it, and no document available on 2026-10-04 answers it.