zhudong.si
DEEP RESEARCH REPORT

XAIxAI / SpaceXAI Deep Research Report

Persistent Agents, AI R&D Feedback, and Compute Systems: From Model Lab to Intelligence Infrastructure

Research cutoff:Oct 6, 2026Chapters:19Full text:~66k wordsVersion:V1.0

Persistent Agents, AI R&D Feedback, and the Compute System: From Model Lab to Intelligent Infrastructure Organization

Research cutoff: October 7, 2026 | Version: GPT V1.0 | Independent company study

This report re-establishes its judgments from current public sources. Its research subject is xAI and the model, Agent, and compute businesses currently operating under the SpaceXAI brand, along with their relationship to SpaceX and Cursor. No conclusions from other company studies are taken as premises.

The report answers three questions: first, through which mechanisms xAI's capability growth is occurring; second, whether persistent Agents and R&D tooling have begun to form feedback that accelerates the growth of intelligence; and third, whether hyperscale compute and space resources can constitute a sustainable advantage.

Reading conventions. A = independent evaluations, original government documents, or external research; B = first-hand company disclosures; C = authoritative media; D = this report's analysis. Preprints in A are separately marked "preliminary" and do not equal peer review or independent replication. Confidence levels express this report's confidence in specific judgments, not the probability of ASI. Citations use [Sxx] source numbers, with accessible links, dates, and usage scope provided at the end. All dynamic evaluation figures are snapshots as of the cutoff date.


Executive Summary

This report's core judgment: the change in xAI most worth studying is that models, persistent Agents, development workflows, and compute supply have begun entering a single organizational system. Its research value already extends beyond Grok's one-off model rankings. But public evidence does not yet prove this system has translated into autonomous, continuous AI R&D takeoff. [D, medium-high confidence]

Eight judgments form the report's main thread:

  1. Organizational change has occurred; capability synergy still needs verification. xAI joining SpaceX and Cursor joining SpaceX both have official records. Acquisitions bring resources and control rights; whether they shorten R&D cycles and improve experiment quality is a separate question that requires data to answer. [S01–S03]
  2. Grok's growth already contains an AI→AI link. The 4.7 model card discloses using 4.6 to generate training trajectories and using model checks to filter out problematic data. This is official evidence of a previous generation participating in the next generation's training; it is insufficient to prove that research planning, experiment selection, and cross-generational growth have autonomously closed the loop. [S05]
  3. Persistent Agents are first a runtime and organizational mechanism. Identity, memory, event triggers, cloud work environments, and tool permissions let work continue across sessions. That alone does not prove the base model is continuously learning, much less that the system already possesses human-employee-like responsibility capacity. [S08–S11]
  4. Multi-agent scale is not a sufficient condition for collective intelligence. Workflow budgets and internal engineering cases show orchestration has entered products. Whether it exceeds large human organizations depends on task decomposition, verification, resource contention, error correlation, and coordination costs. [S07, S09]
  5. Independent evaluations support progress while revealing cost constraints. Grok Build rose from 47 (4.6 combination) to 56 (4.7 combination) on Artificial Analysis's current same-version Agent index; average API cost and run time per task rose accordingly. Success rates alone cannot demonstrate efficiency gains. [S15]
  6. Compute advantage and model advantage must be calculated separately. Colossus providing access to Anthropic is a company-official disclosure. Infrastructure can serve competitors; total device counts cannot be directly equated with training resources exclusively available to Grok. [S17]
  7. Orbital compute is a strategic option requiring engineering validation. FCC application acceptance, customer expressions of interest, and future test plans belong respectively to regulatory procedure, commercial intent, and roadmap planning; none counts as proof of GW-scale orbital operation. [S20, S22]
  8. Control must expand alongside runtime systems. Current Bot product documentation states that one user's Bots share a cloud computer, files, and login state. Role division does not automatically produce permission isolation. Beyond model refusal evaluations, cross-Bot authorization, memory auditing, abort propagation, and transaction verification are needed. [S11, S23]

The baseline judgment for the next 3–10 years is not an ASI year but three turnarounds to test: whether persistent Agents can diffuse from internal cases into repeatable enterprise working practices; whether AI R&D shifts from local assistance to causally verified growth acceleration; and whether industrial resource expansion can continue under cost, safety, and social constraints.


1. Identify the Research Subject First: Names, Control Rights, and Businesses Must Not Be Conflated

1.1 Reliable Organizational Facts as of the Cutoff Date

On February 2, 2026, xAI officially announced joining SpaceX. Cursor officially announced joining SpaceX on August 14; its statement also said the collaboration between the two sides began in April. July's Grok 4.5 launch materials already mentioned collaboration with Cursor. Therefore, not all joint R&D results can be attributed to the acquisition completed in August. [S01–S03, S06, B]

The Grok 4.7 model card lists SpaceXAI as the DBA name of XAI LLC. Other official website pages use SpaceXAI LLC in their footers. This report adopts "xAI/SpaceXAI" as the research name and does not reconstruct all legal entities, equity, and asset ownership from a brand, a footer, or a single model card. The specific legal entity should be determined by the relevant contracts and registrations. [S05, S11, B]

This report's judgment [D]: It can currently be confirmed that business and brand integration has occurred; these materials cannot be used to directly infer that all assets are already within one legal entity, or that all data, accounts, and technologies are unconditionally interoperable.

1.2 Three Businesses Have Different Success Criteria

Business Core output Success criteria to verify Most likely misreading
Frontier model R&D Base capabilities for reasoning, coding, and tool use Reliable capability at equal budgets, training efficiency, cross-task transfer The cluster is large, so the model must lead
Agent and development platform Sustainably executed work, code, and organizational memory Full-work delivery rate, human input, error recovery, net cost Many Bots, so organizational capability already exceeds humans
Compute supply Usable and stable computing services Effective compute, delivery time, utilization, unit cost, and cash recovery Compute is rented out, so the in-house models failed or all devices are idle

These three businesses can reinforce each other, or compete for the same resources. Model training demands R&D priority, customer inference demands stable delivery, and Agents demand large volumes of interactive runtime. Resource allocation conflicts may become a real bottleneck after organizational integration.

1.3 Don't Treat Founder Narratives as Technical Evidence

Public positioning such as "pursuing truth," "building at scale," and "long-running AI colleagues" can explain strategic direction; it cannot prove models are more truthful, R&D more effective, or employees already replaced. Founder visions, product designs, and technical results should be recorded separately.

As of this report's cutoff date, no independently verifiable company-wide AI R&D automation ratio, model R&D cycle times, training failure rates, or a list of algorithm innovations independently proposed and validated by AI has been obtained. These unknowns cannot be filled in by management style or historical execution capability.


2. Evidence of Capability Growth: Progress, with Cost and Caliber Issues

2.1 Current Technical Changes in Models

September 21's Grok 4.7 launch materials state that, compared with 4.6, it uses a larger base model, longer RL training, and higher weighting on difficult, multi-hour tasks, while natively training understanding of the Grok Bot runtime framework. Official starting API prices are $2 per million input tokens and $6 per million output tokens; fast-track acceleration and pricing are counted separately. [S04, B]

This provides a testable growth mechanism: expanded base capability + long-task training + runtime-framework adaptation. It does not prove how much each of the three factors contributed. Larger models, different inference budgets, and better runtime environments may all affect results simultaneously.

This report does not assert "first overall" from launch tables. Identifying genuine technical gains requires cross experiments: fix the model and swap the framework; fix the framework and swap the model; compare inference tiers at fixed cost; test repeated runs on fixed tasks. The company has not disclosed complete ablation results for these factors.

2.2 The Capability–Cost–Time Relationship from Independent Evaluations

Artificial Analysis's cutoff-date page records Grok 4.7 xhigh's Intelligence Index v4.3.2 at 46, with an average evaluation task cost of $3.74 for that index. That cost belongs to the specified evaluation tasks, not the total cost of completing enterprise work. [S14, A]

Its Coding Agent Index v1.5 records Grok Build as follows: [S15, A]

Model combination on same Agent framework Agent index Avg API cost / task Avg run time / task
Grok Build + Grok 4.6 xhigh 47 $3.57 19.5 min
Grok Build + Grok 4.7 xhigh 56 $8.82 39.2 min

This report's calculation [D]: The index rose 9 points, recorded API cost was about 2.47× the former, and run time about 2.01×. The index is not a linear utility function; "56÷47" cannot be used to explain intelligence growth in percentage terms. This table supports capability improvement and resource consumption increasing together under that evaluation caliber; it does not support necessarily higher total productivity, nor a uniform price for all tasks.

If higher capability solved tasks that were previously completely unsolvable, the cost increase may be worth it; if it merely retries the same simple tasks repeatedly, the added investment may have no commercial meaning. Task value needs to enter the judgment.

"Same framework" here refers to the Grok Build product; it does not mean all underlying runtime versions, inference processes, and service conditions were strictly fixed. The generational table is an independent evaluation snapshot and still cannot serve as a pure causal experiment on a training change.

The official launch page's DeepSWE is 71.0%, labeled high; the current AA combination page records 73% using xhigh, evaluating a specified Agent combination. The two are not the same investment and system settings. Official Terminal-Bench and AA records also differ; this report keeps them separately, does not average them, and does not overwrite one with the higher value. Business decisions use the independent Agent page's complete capability, cost, and time caliber; technical attribution still requires further experiments. [S04, S15]

2.3 Long Runtimes Don't Equal Long-Task Reliability

Proximal's FrontierSWE V2 page specifies 34 tasks, 5 trials per task per model, and a 20-hour budget per trial; Grok 4.7's current mean@5 is 29.5%, with an average cost of $318.77 per trial. The page states scores use task rewards, which cannot be equated with full-task completion rates; its error displays also cannot be called statistical confidence intervals without methodological verification. [S16, A]

METR's Task Horizon uses the time human experts need to complete a task as the horizontal axis, fits AI success probability, and finds the time boundary for 50% or 80% probability. Tasks are mainly verifiable software, machine-learning, and security work, which cannot be directly extrapolated to all social activity. [S13, A]

Therefore, these four quantities should be separated:

Quantity Question it answers What cannot be inferred
20-hour runtime budget How long the system can spend at most It can complete all human 20-hour tasks
12 hours actually run How long this Agent run lasted 12 hours of unsupervised reliable work
29.5% partial reward Average reward earned under the scoring criteria 29.5% of tasks fully completed
80% Task Horizon Which human-task duration corresponds to 80% success probability Organizational reliability of continuous weekly operation

This study did not obtain an independently measured METR 50%/80% time boundary directly applicable to Grok 4.7, and does not fill in estimated hours. The absence of that number also does not mean no related capability exists or that the lab did not provide access to evaluation institutions.

2.4 Which Real Metrics Are More Useful Than Rankings

For the xAI route, prioritize measuring four outcomes: work quality when one Bot continuously holds the same responsibility; correctness of cross-system operations; the proportion of proactive error discovery and correction in unfamiliar problems; and the total human time required to complete a delivery.

The same work should be divided at least into "completed on first attempt," "completed after review," "completed after rework," and "not completed." Only by including rework, supervision, and omissions can one judge whether agency has translated into economic capability.


3. Persistent Agents: From Session Growth to Responsibility Growth

3.1 Product Mechanisms Already More Complete Than a Chat Window

Grok Bot's design article places identity, memory, runtime environment, tools, routine tasks, and outputs within a system of persistent work. Work can be activated by prompts, schedules, or events; different roles can be organized by a coordinating Bot. Grok Build memory records cross-session information in Markdown topics at project and global scope. [S08, S10, B]

This report's judgment [D]: Persistence first changes the work container. A model can stop after one inference, but tasks, memory, files, and event triggers persist. On next activation, the system rebuilds state and continues.

This mechanism makes "research materials today, follow up tomorrow, verify next week" an engineering-feasible process, without assuming the model thinks continuously at all times. External memory can improve specific work performance, but alone it does not prove parameter updates or global cognitive improvement.

3.2 How the Runtime System Forms a Closed Loop

The following is an analytical diagram built by this report from public designs, not the company's actual backend architecture:

flowchart TD
    A["Tasks & events"] --> B["Responsibilities, memory & permissions"]
    B --> C["Model planning & tool execution"]
    C --> D["Independent verification & human decisions"]
    D --> E["Outputs, logs & experience updates"]
    E --> B
    D --> F["Abort or rollback"]

Where the system exceeds a single model response lies in its ability to maintain context, access real resources, and update subsequent actions with execution results. Yet the system's reliability ceiling may be determined by its weakest link: a wrong account, expired memory, an invalid verifier, or an un-rollbackable write operation can all cause a highly capable model to fail delivery.

3.3 The Most Important Boundary: Roles Don't Equal Security Isolation

The current Grok Bot FAQ states that one user's Bots share a persistent cloud computer, including files, browser, and login state; isolation is at the user level, not the Bot level. The FAQ also says an enterprise version is available, sensitive operations "can" go through Auto Review, and enterprise network and data control options are provided. [S11, B]

An earlier design article used the interface phrasing "each Bot has its own computer." For the current isolation mechanism, this report prefers the more direct product FAQ; it does not treat role interfaces as independent virtual machines or independent credential boundaries.

This report's analysis [D]: One user can have finance Bots, procurement Bots, and research Bots collaborate, but their task names cannot substitute for technical permissions. Under a shared browser login, a role that should only view materials may touch another role's business accounts. Whether execution is actually permitted requires backend authorization and action policies, not observation of avatars or role descriptions.

This also constitutes a counter-effect of multi-agent development: resource sharing lowers handoff costs while potentially expanding failure blast radius. Multiple Agents editing the same file, operating the same shopping cart, or competing for the same page state produce organizational-style concurrency problems.

3.4 Team Bots Add Organizational Context

The September 28 Team Bots launch materials state that they share skills and related work materials, while retaining each user's private session context and memory. An engineering case discloses Bot-coordinated Cursor Projects and hundreds of Cloud Agents helping a 5-person team deliver over 100 PRs per day and ship Team Bots within weeks. [S07, B]

The case shows AI participating in building AI products, with orchestration entering real engineering. Still missing: PR size, merge definitions, defect rates, human review minutes, control teams, and long-term maintenance costs. 100 PRs may contain many small changes; they cannot be directly converted into 100 independent research results.

Shared skills, private context, and shared organizational knowledge also have a boundary needing verification: how is derived knowledge learned from private sessions labeled and permitted to be shared? Public descriptions are insufficient to prove leakage exists, and insufficient to prove all derived-information isolation has been resolved.

3.5 What One Internal Customer-Support Case Can Demonstrate

The company's September 22 disclosure reports ticket volume grew 175% without adding headcount, and reports usage costs as low as $0.20–$0.30 for some optimization tasks. Implementation started from internal suggestions with human approval for every write operation, then added run logs, evaluations, and direct responses for simpler tickets. [S12, B]

175% growth means ticket volume became 2.75× the original, not 1.75×. The low usage costs exclude all integration, supervision, subscription, error compensation, and organizational change costs; "how many people might have been hired" is the company's counterfactual estimate.

This report's judgment [D]: Compared with "AI fully takes over" marketing language, gradual authorization and iterative feedback more closely resemble the engineering process by which capabilities actually form. Only if third parties can reproduce such quality and cost results can persistent Agents become an organizational productivity metric.


4. AI R&D: Feedback Exists, but Takeoff Needs Stronger Evidence

4.1 Decomposing "AI Improves AI" into Five Levels

Level Object improved xAI public evidence status This report's assessment
L1 Researcher tools Code, retrieval, debugging, documentation Development products and internal engineering cases Engineering usable; company-wide R&D benefit unknown
L2 Agent runtime systems Workflows, memory, coordination, and product code Workflows improvable across runs; Team Bots engineering case Local feedback with official evidence
L3 Model training inputs Trajectories, synthetic data, filtering signals 4.6 participated in 4.7 training-data generation and checking AI→AI official evidence exists
L4 Research and training loop Propose hypotheses, allocate compute, train, evaluate, select new algorithms No complete public run records obtained Autonomous closed loop cannot be confirmed
L5 Cross-generational growth acceleration New generations improve next-round R&D capability and accelerate continuously No independent, budget-controlled cross-generational evidence obtained Takeoff cannot be confirmed

L2 per [S07, S09], L3 per [S05], all B. L4 and L5 are this report's evidence-gap conclusions, not "the lab internally must not have done it."

4.2 AI→AI Chains That Have Already Appeared

The 4.7 model card discloses that 4.6 generated SFT trajectories with model checks filtering them; supplementary training also used anonymized Cursor workflow data. [S05, B]

This already goes beyond "only helping programmers write code." It lets the previous generation enter the next generation's training production process. The key unknowns: data proportion, model screening accuracy, degree of human correction, that link's contribution, and its substitution rate for other sources.

This report's judgment [D, medium-high confidence]: The publicly disclosed xAI roadmap contains components of R&D feedback; whether these components have formed continuous positive feedback that raises the growth rate of intelligence is currently unknown.

4.3 The Difference Between Two Kinds of Feedback

Product feedback is Bots generating code, evaluations, memory, or workflows that make the same base model work better. It may improve many customer tasks without necessarily changing base capability.

Research feedback is AI proposing or screening new methods so that next-generation training achieves higher generalization at fixed resources, making subsequent research more effective. To prove it, one should trace the origin of research suggestions, experiment execution, independent validation, adoption decisions, and next-round results.

The two can reinforce each other: better Agents may organize more experiments; better models may generate higher-quality research suggestions. They can also disconnect: software ships extremely fast while key algorithm research still depends on a few researchers; synthetic data increases while generated errors and reward biases are also amplified.

4.4 Separate "Continuous Shipping" from "R&D Acceleration"

Shorter release intervals can come from parallel training, renaming existing models, product packaging, more capital investment, or release strategy changes. The simultaneous appearance of larger base models and longer training also cannot be used to infer research efficiency from short release intervals.

Verifying AI R&D acceleration requires at least four sequences:

If only AI execution volume grows while researchers' waiting time for training results stays unchanged, local automation does not equal overall takeoff.

4.5 Recursive Self-Improvement Can't Be Blanket-Denied, Nor Transplanted as Company Fact

The September 22 AIDE² preprint reports a research Agent undergoing 7 consecutive self-code improvements during an 8-day run, with transfer gains across 4 held-out evaluations. That evidence belongs to external research, author reports, and limited experiments; this report verified the abstract but did not reproduce the full-text experiments. [S26, A—preliminary]

It shows "recursive improvement at the Agent code level" has become a measurable engineering object. It does not prove xAI uses that system, does not prove base-model weights improve autonomously, and does not prove long-term gains offset diminishing R&D returns.

Object-explicit language should be used: workflow self-revision, memory updates, recursive optimization of Agent source code, training-data feedback, and base-model R&D loops. Calling all of them "self-evolution" makes the most important changes unmeasurable.


5. The Object of Scaling Has Expanded: Adding Resources ≠ Adding Effective Intelligence

5.1 Which Expansions xAI's Current Public Roadmap Contains

Dimension Publicly disclosed mechanism Expected benefit [D] Limit to test [D]
Pre-training Larger base model [S04] Broader base representations Data quality, compute budgets, marginal returns
Supplementary training Model trajectories and Cursor workflow data [S05] Closer to technical and real work Generation bias, authorization, data selection
Post-training & RL Longer RL, difficult tasks, Agent environments [S04] More reliable execution and self-checking Reward sparsity, verifier exploits, training distribution
Inference compute Different inference budgets; independent cost records [S14–S15] Multi-step handling of hard problems Cost, latency, and repeated inference
Context & memory 500k context; topic memory [S10, S14] Continuing long work and experience Attention quality, expired memory, recall
Agent runtime Persistent environments, event triggers [S08, S11] Continuing action across sessions Resource contention, state corruption, unsupervised operation
Parallel Agents Orchestratable workflows [S09] Parallel processing of decomposable tasks Communication and merge costs, correlated errors
Verifier Workflow verification and model data checks [S05, S09] Filtering low-quality results Same-origin bias, reward hacking, insufficient coverage
Tools & resources Software, browsers, enterprise credentials [S07, S11] Turning suggestions into work outcomes Permissions and rollbackability
Industrial compute Colossus self-use and external access [S17–S18] More training and runtime supply Available power, effective interconnect, economic returns

The mechanisms in the table are not ten independently quantified scaling laws. The company has not disclosed sufficient controlled experiments for this report to fit a unified growth law.

5.2 The Relationship Between 2020-Style Scaling and Current Expansion

This report uses "scaling" to mean increasing a class of resources and testing capability gains. The resources under discussion now include not only pre-training parameters, data, and compute, but also execution duration, parallel experiments, tools, memory, feedback, and verification.

These resources often substitute for each other: a better verifier can reduce wasted search; better memory can reduce repeated retrieval; more tools may reduce what the model must memorize; longer context may raise costs without improving key-fact usage quality.

Therefore, the question should be "for which task, at fixed which budget, increasing which resource, how do results change" — not merging all growth into one scale metric.

5.3 A Framework for Attribution

This report proposes the following conceptual relationship, not an empirical law:

Effective work output = successful deliveries × value per delivery − human supervision, errors, coordination, and resource costs.

A higher capability score changes only one of these conditions. If work requires repeated rework, human review effort rises in tandem, or most of the cluster serves failed-result verification, compute growth cannot directly convert into economic growth.

One can further define R&D conversion rate: the increment of new capability validated on held-out tasks per unit of complete R&D investment. It is closer to the mechanism this report studies than "GPUs/headcount." No xAI public data currently available can compute this metric.


6. Systemic Superintelligence: Three Competing Explanations

6.1 Hypothesis H1: A Strong Central Model Determines the Ceiling

If the base model cannot recognize research errors, propose valid hypotheses, or understand tool feedback, adding Agents may merely replicate the same defect. Grok 4.7 still expands the base model and strengthens training, indicating the public roadmap has not abandoned model capability. [S04]

H1 predicts: cross-framework capability highly correlated; swapping in a stronger model gains more than adding more roles; bottlenecks in hard tasks concentrate on judgment and generalization. If these phenomena persist, the model is the key constraint and the system mainly releases capability.

6.2 Hypothesis H2: System Combinations First Form Capabilities Beyond Organizations

Persistent responsibilities, shared knowledge, permissions, asynchronous execution, specialized tools, and verifiers may let model combinations without fully superhuman capability first exceed large human teams in limited domains.

H2 predicts: with the model fixed, system improvements still significantly raise full-work success rates; at equal budgets, large tasks are correctly decomposed and merged; new employees or new Bots inherit organizational experience faster; errors don't accumulate over long operation.

xAI's public product layout is compatible with H2, but current internal cases have not yet completed same-task, equal-cost human-organization controls. [S07–S12]

6.3 Hypothesis H3: Industrial Infrastructure Is the Main Sustainable Advantage

The third explanation: model and Agent competition changes fast, while low-cost, fast-delivered, reliable compute services are harder to replicate. Even if Grok doesn't lead in all capabilities, infrastructure can still serve other labs.

H3 predicts: external compute customers grow, revenue decouples from in-house model rankings; deliverable compute capacity grows ahead of exclusive model demand; business focus moves closer to service stability, energy, and capital management. [S17–S19]

6.4 Current Choices and Possible Counter-Evidence

This report's judgment [D, medium confidence]: H1, H2, and H3 can all hold simultaneously at present. The more explanatory framing is not "single model versus system" but: models determine judgment quality, systems determine actual work scope, and infrastructure determines extensibility.

If system redesign yields no sustained gains with the model fixed, H2's weight should be lowered; if effective compute resources keep increasing while held-out capability stagnates, the weight of "compute directly converts to model advantage" should be lowered; if external customer contracts and infrastructure returns cannot be sustained, H3 cannot be treated as a certain fallback either.

6.5 Multi-Agent Is an Optional Organizational Form

Independently handleable code reviews, information retrieval, and experiment branches suit multi-agent; work requiring unified state, continuous action, or hard-to-split judgment may better suit single-Agent with tools and strict verification.

Official workflows give a typical budget of 128 Agents, up to 1,024 for larger tasks. The budget ceiling is not actual concurrency, much less the capability sum of 1,024 independent agents. [S09, B]

This report suggests controls at scales like 1, 4, 16, and 64 with the same total cost, recording actual simultaneous activity, effective independent results, conflict rates, and integration time. This is a study-design example, not a company-published experiment.

A system can only be called organizationally superhuman in a domain when it stably exceeds the corresponding human organization on full-process work, delivery time, and total cost. Generating large volumes of text, PRs, or experiments within hours is still insufficient for this standard.


7. Colossus and Capital: Device Inventory Must Become Usable Compute

7.1 Three Sets of Numbers Cannot Be Directly Added

Official material Numbers and time caliber How this report uses them
Colossus intro page Top line 200,000 H100s; another metric on the same page 180,000 H100s; historical narrative of 122 days to build, 92 days to expand; also a million-GPU roadmap [S18] For understanding historical build-out; cannot be taken as cutoff-date complete inventory
Jan 6, 2026 financing disclosure End-2025 Colossus I and II combined exceed 1 million H100-GPU-equivalent compute [S19] Recorded as the company's equivalent-compute caliber, not converted to physical GPU counts
May 6, 2026 partnership disclosure Colossus 1 exceeds 220,000 NVIDIA GPUs, including H100, H200, and GB200 [S17] Uses the newer, more scope-defined company disclosure; does not extrapolate group-wide totals

"Million equivalent compute," "million physical devices," "one interconnected cluster," and "all group data centers" are four different concepts. Mixed models, precision, bandwidth, utilization, and workloads affect equivalent conversion. This report did not obtain the complete definition of that conversion, nor an audited cutoff-date device inventory.

Different counts on the same page are unsuitable for elimination by averaging. The proper treatment is to keep the version, scope, and units, lowering confidence in real-time inventory. The website's "world's largest" phrasing is a B-grade company claim; this report has not completed an independent comparison of all comparable global clusters.

7.2 The Meaning of Supplying Compute to Anthropic

The May 6 official statement says an agreement was signed to provide Anthropic with Colossus 1 access, with Anthropic planning to improve Claude subscription service capacity. The statement also mentions Anthropic's expressed interest in multi-GW orbital compute collaboration. [S17, B]

It confirms "a competing lab can also be a compute customer." It does not disclose that all group devices have been leased out, specific utilization rates, exclusivity terms, customer revenue, or that orbital collaboration has become a definitive construction contract.

This report's analysis [D]: External customers can raise asset utilization and revenue, and may also occupy opportunity resources for in-house models. The company gains scale and cash inflow; customers gain usable compute; both sides may benefit. Without contract and resource-allocation data, the leasing action cannot be used to judge whether the model roadmap succeeded or failed.

7.3 Compute Services Face at Least Six Bottlenecks

  1. Chips and HBM. Device counts are constrained by advanced-chip, memory, and packaging supply; chip arrival does not mean all supporting components are complete.
  2. Interconnect. Connectable topology, effective bandwidth, failure recovery, and communication overhead determine how many devices can train together.
  3. Stable power. Rated power, sustainable supply, load-fluctuation handling, and grid-connection timing should be recorded separately.
  4. Cooling. High power density affects rack design, cooling equipment, construction timelines, and outage risk.
  5. Compute demand and scheduling. Training, inference, external leasing, verification, and experiments cannot double-count the same capacity.
  6. Capital recovery. Device depreciation, financing, energy, and customer payments determine whether assets can keep expanding.

These belong to this report's industrial analysis. The company has not disclosed all variables; it is inadvisable to estimate total cost of ownership or current effective training FLOPs.

7.4 Power Is the Industry's Hard Constraint; Global Ratios Alone Won't Do

The IEA's current report projects global data-center electricity consumption rising from 485 TWh in 2025 to 950 TWh in 2030, about 3% of global electricity demand; this is a central-scenario forecast, not an already-realized result. The report notes bottlenecks in HBM, grid connection, and energy equipment, and that on-site gas generation is not necessarily faster to deliver than the grid. [S24, A]

A limited global share can coexist with a single city bearing large new loads. For Colossus, what matters more is local available capacity, substation and gas utility support, operational volatility, community impact, and permit conditions — not converting global electricity into "how many more clusters could theoretically be built."

7.5 External Environmental Measurements Should Enter the Evidence Map

Gauld et al.'s NOx study preprint, calibrated with TEMPO satellite observations and plant continuous-monitoring data, estimates average emissions of 730±185 kg/hour after February 2026 from power plants associated with Colossus 2 — about 16× the value derived from permit and control-technology assumptions. The abstract of the paper's September 15 revised version has been verified. [S25, A—preliminary]

This is a research estimate with modeling and measurement uncertainty, not this report's own measurement; it does not equal a final enforcement determination of violation, nor does it directly prove any resident health outcome. Complete methodological review, ground monitoring, and facility operations information are still needed for cross-verification.

It nonetheless changes the research question: the benefits of expansion speed must be evaluated together with pollution control, on-site energy reliability, community acceptance, and future costs. Infrastructure management is not just about connecting GPUs.

7.6 Capital Scale ≠ Sustained Funding Availability

The official January financing disclosure was a $20 billion Series E. [S19, B] That is financing on a definite date, not October cash balances, annual AI R&D spending, data-center asset net value, or model training cost.

This report does not judge current investment value by historical transaction valuations, nor has it obtained audited financials separating the model, Agent, leasing, and space businesses. For future paths, monitor contract continuity, capital costs, effective utilization, refresh cycles, and business cash recovery.

This report's judgment [D]: AI in the xAI roadmap already involves industrial systems. Software feedback may change daily, while grid connections, construction, supply chains, and permits change on monthly or longer cycles. The so-called R&D takeoff may first collide with resource delivery speed.


8. Orbital Compute: Transferring Constraints, Not Removing Them

8.1 Levels of Publicly Available Evidence

On February 4, the FCC published its acceptance and comment-solicitation notice, confirming SpaceX's January 30 application for an orbital data-center system of up to 1 million satellites, involving 500–2,000 km altitudes and optical inter-satellite links. [S20, A—government procedure fact]

Acceptance ≠ approval; approval ≠ launch; launch ≠ stable compute; stable compute ≠ cheaper than ground training. This study did not obtain the system's complete final authorization and commercial capacity proof as of the cutoff date, and cannot use the February notice alone to assert the approval's current final status.

Reuters reported in June that SpaceX targets orbital AI compute tests before the end of 2027; that reporting depends on its obtained information and sources and should be classed as C-grade plans. Early-October media reported industry prototype payload tests; prototype-level progress still has distance from GW-scale commercial deployment. [S22, S29, C]

8.2 Why This Choice Has Strategic Value

Supporting reasons [D]: Space resources may bring launch, batch manufacturing, communications, and operations synergies; certain orbits have high solar-utilization potential; some compute can be placed near space data sources, reducing backhaul; local land, grid, and water constraints on the ground may make alternatives more attractive.

Opposing reasons [D]: Ground bottlenecks are converted into payload mass, reliability, radiation, thermal control, link, maintenance, and orbital governance problems. Launch advantages may not cover the costs of short computing-equipment lifespans, rapid refreshes, and high-frequency servicing.

Judge by workload. Tasks that are independent, delay-tolerant, and already have data in orbit are not the same compute market as huge model training requiring high-bandwidth synchronization.

8.3 Vacuum Is Not a Free Cooling System

NASA thermal-control materials explain that in vacuum there is no convective heat transfer; satellite internals mainly transfer heat by conduction and exchange it externally mainly by thermal radiation; solar, Earth-reflected, and infrared inputs plus material performance degradation must also be considered. [S21, A]

This report makes an ideal-order-of-magnitude calculation [D] solely to illustrate engineering scale, not representing SpaceX designs: to reject 1 GW of waste heat, taking emissivity 0.9, ignoring external heat inputs and background radiation, using

[ A=\frac{Q}{\epsilon\sigma T^4} ]

where the Stefan–Boltzmann constant (\sigma\approx5.67\times10^{-8}\ \mathrm{W\,m^{-2}\,K^{-4}}).

Radiating surface temperature assumption Ideal heat flux Required effective radiating area
300 K ~413 W/m² ~2.42 km²
400 K ~1,306 W/m² ~0.77 km²

Area means effective emitting area, not satellite projected area or ground footprint. 400 K is about 127°C; ordinary computing hardware cannot be directly assumed operable at that temperature; thermal transport and materials solutions are needed. The formula excludes solar panels, structures, shading, energy storage, links, mass, and servicing, and cannot be used to estimate total cost.

The calculation's purpose is problem identification: at GW scale, thermal management remains large-scale systems engineering; space's low background temperature does not automatically carry waste heat away.

8.4 Economics Should Be Priced on "Effective Compute"

Orbital options should compare: full-lifecycle available compute, multiplied by real task efficiency, divided by manufacturing, launch, replenishment, link, operation, and failure costs. Ground options should also fully include energy, grid-connection, and cooling costs.

Evidence needed includes: available power rather than solar-panel nameplate ratings; in-orbit payload error rates; task-level interconnect efficiency; long-term thermal-control data; per-device replenishment cycles; actual service pricing or costs. Announcing satellite counts does not provide these answers.

This report's judgment [D, medium confidence]: Orbital compute is worth studying as an option during 2026–2036, but is currently insufficient as a basis assumption for xAI model leadership or near-term ASI onset. The clearer existing assets remain ground compute and digital work platforms.


9. World Models and Physical Intelligence: Engineering Software Capability Hasn't Crossed Real-World Validation

9.1 Technical Scope That Can Be Confirmed

Grok's current disclosures include engineering, CAD, and compute-task capabilities and corresponding training environments. [S04–S05] They support the judgment "AI enters industrial design software." Available verification materials are insufficient to prove xAI has publicly completed full validation of general physical world models, VLAs, or robot foundation models.

No public sufficient evidence differs from no internal research — they are different propositions. This report does not invent undisclosed roadmaps, nor treats video generation effects, CAD scores, or having a space business as substitute indicators of robot reliability.

9.2 Three Different Physical-Capability Chains

Capability chain What should be verified Why real-world validation is still needed
Engineering design Design constraints, simulation, manufacturability Simulations may miss materials, tolerances, and failure conditions
Physical prediction Dynamic consistency, counterfactuals, state estimation Coherent-looking visuals may not support control
Embodied execution Perception, action, contact, recovery, safety Real objects, environments, and sensors have distribution shift

Digital results can be copied and quickly evaluated; physical results still need materials, equipment, and real trials. If AI first optimizes compilers, kernels, or experiment scheduling, then helps manufacturing equipment, independent physical constraints still lie between the two stages.

9.3 The Tesla Connection Doesn't Automatically Fill the Evidence Gap

January official materials mention Grok Voice's use in Tesla vehicles. [S19, B] That proves a product connection, not that xAI, SpaceX, and Tesla's robot training, property rights, data, and R&D have all merged. Common founders or visible product connections cannot replace concrete technical interfaces and experiment records.

This report did not obtain sufficient materials to verify a cross-company "digital research → robot control → real-world experiments → model improvement" loop, and does not write it as an established fact.

9.4 Digital and Physical May Have a Time Lag

Scenario judgment [D]: In software and some compute-research domains that are automatically verifiable and parallel-testable, organizationally superhuman capability may appear first; real robots, production lines, and space manufacturing need longer validation cycles.

This time lag cannot be pre-filled as three or five years. If automated experimental equipment, low-cost robot deployment, and high-fidelity simulation mature simultaneously, the gap may shorten; if real-world accidents, hardware reliability, or certification become bottlenecks, it may widen.

Quarterly monitoring should record both simulation and real results, disclosing task distribution, repetition counts, human interventions, safety stops, and cross-environment transfer. A single successful demo does not provide these quantities.


10. Control and Governance: From Refusing Dangerous Answers to Constraining Real Actions

10.1 The Current Framework Can't Be Judged by Old Impressions

The Frontier Artificial Intelligence Framework published June 30 covers four risk categories: CBRN, offensive cyber, loss of control, and harmful manipulation. The document proposes at least one complete system risk assessment per year, and lists additional assessment triggers including model updates, serious incidents, and increased integration risk; development and deployment rest on the company's judgment that risks are acceptable. [S23, B]

Thus "xAI has no safety framework" or "only looks at malicious use, not loss of control" does not match this current document. Policy commitments also don't equal independent verification: internal execution, exceptions, veto-power use, and final risk judgments still need evidence.

This study did not find directly computable mandatory numeric thresholds specifically targeting AI R&D automation in the public framework. Model cards' biological, security refusal, and other results can illustrate the tested scope; alone they cannot prove the overall controllability of persistent Agent organizations. [S05, S23]

10.2 Four Safety Objects Must Be Separated

Object Main question Meaningful evidence
Model outputs Whether they provide dangerous assistance, deception, or wrong high-confidence answers Independent attack tests, scope, versions, and failure samples
Agent execution Whether it exceeds authority, makes wrong transfers, wrong deployments, or bypasses approvals Action logs, permission denials, approval accuracy, and rollbacks
Agent organizations Whether errors propagate across Bots and supervision fails Cross-Bot attack tests, abort propagation, shared-state audits
R&D and infrastructure Whether training can be stopped, weights protected, and key resources limited Real drills, independent reviews, shutdown verification, and accountability records

The stronger the model and the more tools, the more risk comes not just from "what it knows" but "what it can do under what permissions." Permissions chosen by the company or customer constitute part of system capability.

10.3 Shared Environments Need Independent Action Controls

This report proposes five Control mechanisms to verify [D]: tiering read, draft, write, publish, and pay actions; binding authorization to actions and objects; preventing untrusted-source web pages from changing higher-level permissions; ensuring one Bot's abort cancels derived tasks; and retaining checkable outcomes and responsible parties for real external actions.

When the same model proposes and approves, errors may correlate. Adding a "review Bot" does not automatically produce independent review, especially when it shares the same memory, tool state, and reward objectives. Sensible controls should include permission rules outside the model and verifiable outcomes.

10.4 Deployment Safety Already Has Real Disputes

The European Commission's January 26 notice launched a DSA investigation into X, assessing Grok feature deployment risks and related risk assessments. The UK ICO launched a data-processing investigation involving XIUC and X.AI on February 3, concerning systems that could generate harmful sexualized images and videos. Both notices are investigation facts, not final conclusions of illegality. [S27–S28, A—government]

These events differ from ASI loss of control, but provide real evidence on the same governance question: product releases, public distribution, and personal data protection cannot rely on model-level testing or vision alone. Follow-up formal decisions, mitigation measures, and actual recurrence rates should be tracked, not just whether policy documents exist.

10.5 "Truth" Needs Measurable Standards

Real-time social information is timelier, but may be full of manipulation, noise, and citation cycles. A system that writes news, discussions, and synthetic content back into memory creates a recursion problem for factual sources.

This report suggests measuring factual accuracy, confidence calibration, source traceability, correction latency, and consistency on value-sensitive questions separately. Different value orientations can be chosen, but value positioning cannot substitute for factual quality and source auditing.


11. Benefits and Conflicts of Platform Integration: Who Should Superintelligence Serve

11.1 Cursor Brings the Work Entry Point, and Changes Supply Relationships

This report's analysis [D]: A development work platform can provide long-term project state, execution feedback, enterprise procurement entry points, and authorizable data. Compared with selling just an API, it has a better chance of forming a "use → feedback → training or runtime improvement → continued use" cycle.

But tension exists between the in-house model and the multi-model platform: the platform wants the models most beneficial to users; the model business wants traffic and data; the two businesses' commercial objectives may not align.

OpenAI stated on August 28 that it intends to terminate its model-supply contract with Cursor, proposing November 12 as the deactivation date, citing concerns about control changes and contract compliance. That content is OpenAI's one-sided statement; its allegations cannot be written as an independent ruling; the cutoff date also cannot write a future date as a completed deactivation. [S21A, B]

This event suffices to show that a multi-model platform's supply relationships may change with ownership changes. Platform efficiency, supply continuity, and user asset portability all need to enter enterprise choices.

11.2 When Agents Act on Users' Behalf, Commercial Objectives Must Be Checkable

If Bots represent users in procurement, software selection, research, and work assignment, the platform may simultaneously benefit from subscriptions, model usage, compute, or other businesses. An Agent choosing a more expensive runtime, preferring internal products, or retaining more data may not serve users' best interests.

This report does not conclude such potential conflicts are currently occurring manipulation. Verifiable boundaries: whether fees and preferences are disclosed; whether vendors can be specified; whether migration is provided; whether memory can be cleared; whether decision reasons and real operation records are kept.

11.3 Ownership Contains at Least Five Rights

Users should control objectives, memory, credentials, real actions, and migration. The platform owning models and runtime infrastructure should not be automatically interpreted as the platform owning all of users' work context or agency decisions.

Personal Agents, team Agents, and large-group internal Agents have different principals. One shared Bot may simultaneously serve employees, management, and customers; when conflicts arise, it must be clear whose authorization it acts under.

This report's judgment [D]: The social dividing line of system intelligence development may not be "who trains the strongest model" but "who can commission, constrain, verify, and migrate this intelligence." xAI integrating entry points and infrastructure raises the possibility of capability expansion and increases this question's practical importance.


12. Proactive Falsification: Which Judgments Could Be Overturned

This report's provisional judgment Strongest counter-explanation Evidence that would change the conclusion
Persistent Agents are an organizational-mechanism change Just chat plus scheduled tasks; no qualitative capability change Same model, same budget, persistent environments still don't improve full delivery
Model–system synergy deserves attention Core growth mainly from larger models and more inference Cross-framework ablations show small system contribution
Internal cases show engineering feedback PRs subdivided, favorable cases selected, human work hidden Independent review finds delivery or quality net gains disappear
AI→AI components may drive R&D Synthetic trajectories are just routine data engineering Controlled studies find no improvement in effective research cycles or generalization gains
Compute supply may have independent value Customer leasing is short-term gap-filling with insufficient long-term returns External utilization, renewals, or unit revenue keep declining
Orbital compute is currently an option Engineering progress may be faster than public materials suggest Public long-term in-orbit data, real interconnect tasks, and cost advantage
Permission governance is a key bottleneck Stronger models significantly reduce errors and supervision needs Large-scale independent production records show low intervention and low incidents
Digital capability may exceed organizations first Organizational problems insufficiently formalized; physical data is instead more direct Real robot deployments first show broad superhuman results

Three research biases must be avoided in particular.

First, dense product releases manufacture growth illusions. The same base capability can be packaged as multiple features. Track task quality, not news counts.

Second, public-material selectivity. The company knows failed experiments and human workloads; researchers usually see only launch results. Independent evaluations should prioritize hard cases and failure distributions.

Third, tool and process improvements miswritten as intelligence generational growth. For users, the former already has value; for ASI mechanism research, one still needs to identify whether memory, resources, parameters, or research methods changed.

As of the cutoff date, the judgments most needing lowered confidence are: "large-scale Agents already exceed human organizations," "orbital compute will soon remove ground constraints," and "AI R&D takeoff has already occurred." The most certain: visible products and training processes already include persistent execution, Agent orchestration, and previous-generation models participating in next-generation data production.


13. 2026–2036 Scenarios: Turnaround Conditions Instead of an ASI Year

The following scenarios are research tools, allowed to coexist and switch, without pseudo-precise probabilities. The time window is divided into roughly the next 3 years, 3–5 years, and 5–10 years; none is an ASI arrival time.

A | Continued Scaling and Agent Diffusion, No Clear AGI Day

Premises: Models still improve gradually; persistent Agents enter enterprises through clearer workflows; R&D has no confirmable overall takeoff.

Leading indicators: Success rates improve at fixed cost; human review decreases; incident-task reliable delivery; third-party cases replicate across industries.

Outcome: Near term first improves development, customer support, data, and internal operations; medium term stable AI job combinations appear; long term some well-defined domains may form organizationally superhuman capability, while everyday open tasks keep human decisions.

Risks and falsification: Cost and coordination growth eat the gains; long-term memory accumulates errors. If same-model system experiments show no net gain, lower this scenario's weight.

B | AI R&D Shows Verifiable Growth Acceleration

Premises: AI participates not only in execution but in experiment design, algorithm selection, and independent validation; research improvements continue to work on next-round capability.

Leading indicators: Effective experiment output improves at fixed budgets; research cycles shorten; AI-proposed algorithms reproduced externally; gains across two or more consecutive generations have clear attribution.

Outcomes: Agent R&D, data production, kernel and training optimization may accelerate first, later affecting base models. The speed gap between model releases and infrastructure delivery widens; Control needs more frequent re-evaluation.

Risks and falsification: Selection bias, reward exploits, non-independent validation. If all gains can be explained by added compute and human research investment, Takeoff cannot be concluded.

C | Industrial Resources and Governance Form the Main Constraints

Premises: Capability demand growth outpaces chip, power, funding, or permit delivery; complex Agents bring higher operations and supervision costs.

Leading indicators: Queue times increase; announced capacity fails to convert to available capacity; device utilization fluctuates; incidents or external disputes cause delays.

Outcomes: More emphasis on inference efficiency, scheduling, verification, and narrower workflows; infrastructure supply may matter more; model and Agent growth decelerates in phases.

Risks and falsification: Vertical resource integration creates concentrated dependence. If effectively available compute and unit costs keep improving rapidly, weaken this scenario.

D | Technical Paradigm or Per-Resource Efficiency Breakthrough

Premises: Effective new algorithms, memory, or verification mechanisms achieve significant transfer gains at fixed budgets; possibly first proposed by external labs.

Leading indicators: Held-out task results independently reproduced; effective across models and frameworks; not just longer inference; reusable across more tasks.

Outcomes: Relative advantage of existing mega-clusters may be reassessed, or become resources for scaling new methods. Whether open runtime systems, work platforms, and verification assets adapt determines actual gains.

Risks and falsification: Benchmark gaming, narrow tricks called new paradigms. If gains are limited to trained or tuned tasks, don't raise the weight.

E | Compute Supply First Forms an Independent Competitive Advantage

Premises: External labs need capacity long-term; construction, power supply, and scheduling form sustainable unit-cost and delivery advantages.

Leading indicators: Multi-customer renewals, available capacity, and verified unit costs; revenue not fully synchronized with in-house model rankings.

Outcomes: The company may become an important supplier even without leading in all model domains; intra-group compute allocation and customer isolation become governance core.

Risks and falsification: Customer concentration, demand mismatch, high depreciation. Lease counts or contract amounts alone cannot prove profit and sustained advantage.

How to Update Scenarios Quarterly

One news item doesn't trigger a scenario switch. Evidence first changes one variable, then observe whether actual work or research results change. An orbital prototype launch raises that engineering option's maturity; it doesn't directly raise model R&D takeoff probability; a new Agent score raises capability judgments, not directly lowering energy constraints.


14. Leading-Indicator Dashboard: 28 Measurements for Quarterly Updates

Markers: Known = current value obtained; Disclosed = company information, lacking independent verification; Unknown = insufficient public data. Keep each quarter's task sets, model tiers, frameworks, times, costs, and evaluation versions; don't force-connect lines when calibers change.

# Metric Cutoff-date baseline Quarterly update method & notes
1 Cross-domain capability Known: AA Intelligence Index 46 [S14] Compare fixed versions, keep per-domain results
2 Coding Agent capability Known: Grok Build combination index 56 [S15] Same framework, budget, tier, and task set
3 Full delivery rate Unknown: unified enterprise caliber Include failures, rework, abandonment; sampled external review
4 Total cost per successful delivery Unknown; only partial API costs [S15] Include supervision, integration, error, and runtime costs
5 80% Task Horizon No independent 4.7 value obtained Fit using human task durations and success rates
6 50% Task Horizon No independent 4.7 value obtained Alongside 80%, don't cherry-pick the higher one
7 Long-task rewards Known: FrontierSWE V2 mean@5 29.5% [S16] Keep partial-reward definition and trial budgets
8 Human intervention minutes Unknown Human time per successful task, per workday
9 Persistent responsibility reliability Disclosed: persistent Bot products [S08, S11] 7/30-day responsibility tasks, omissions, and remediation
10 Memory effectiveness Disclosed: topic memory [S10] Correct recall, expiry rates, correction and deletion propagation
11 Actual Agent concurrency Unknown; budget ≠ concurrency [S09] Record simultaneous activity and effective independent results
12 Parallel scaling net gain Unknown Fixed-total-cost scale ablations and conflict rates
13 Cross-tool task success rate Unknown Cross-browser, API, and enterprise-software end-to-end measurement
14 AI R&D work share Unknown Recorded by research phase and human hours
15 AI-generated data contribution Disclosed: 4.6 participated in 4.7 data [S05] Quality, quantity, and controlled removal experiments
16 R&D cycle time Unknown Question establishment to reproduced result, including failed training
17 Original methods Unknown: no auditable list obtained Proposal origin, adoption, external reproduction, and budgets
18 Recursive-gain persistence Unknown: insufficient xAI cross-generational evidence Consecutive-generation held-out effects and causal attribution
19 Verifier quality Unknown Misses, false alarms, reward exploits, and independence
20 Effective compute capacity Partially disclosed; scope conflicts [S17–S19] Verify by model, location, availability, and allocation
21 External compute contracts & usage Disclosed: Anthropic access agreement [S17] Actual delivery, renewals, concentration, and resource conflicts
22 Stable power & delivery No full-caliber data obtained Rated/sustainable capacity, outages, and grid-connection delays
23 Environmental & social costs Preliminary: NOx external estimate [S25] Ground and satellite cross-verification, formal decisions
24 Compute economics Unknown Unit effective compute cost, depreciation, and recovery
25 Orbital compute maturity Applications, plans, and industry prototypes [S20, S22, S29] In-orbit power, thermal control, error rates, and task costs
26 Physical task success rate Insufficient company evidence obtained Record simulation and reality separately, repeat across environments
27 Agent controllability Disclosed products and frameworks [S11, S23] Overreach rates, aborts, derivative cancellation, and rollbacks
28 Governance & user control Partial policies and investigation facts [S23, S27–S28] Actual execution, incidents, migration, and data rights

Three signals triggering research-direction adjustment [D]:

"Unknown" should be kept; don't fill it with product counts, release frequency, or founder predictions. This table is for future updates; no automatic monitoring or periodic external tasks were created.


15. What It Means for Enterprises and Individuals: What Assets Can Accumulate

15.1 Enterprises: Choose Verifiable Responsibilities First

Work suitable for pilots usually has clear inputs, readable systems, definite outputs, and verifiable results. Examples: ticket pre-investigation, sales-account fact updates, software issue reproduction, reconciliation discrepancy localization, daily data interpretation.

Phase one has Agents complete preparation and drafts; phase two completes low-risk actions within verified rules; phase three evaluates broader responsibilities. Record quality, failures, human time, and costs at each phase. This phase division is this report's suggestion, not a fixed process all companies must adopt.

Procurement judgments shouldn't compare subscription fees alone. A cheap Agent needing heavy checking may cost more human effort than an expensive but reliable Agent; on the other hand, high-tier inference may also waste on simple work.

15.2 AI2B, FDE, and OPC: Opportunities Come from Organization and Verification

This report's judgment [D]: Persistent Agent platforms may compress differentiation in generic chat interfaces, simple connections, and template processes; the more accumulative value lies in customers' real business definitions, authorization relationships, data quality, evaluation assets, exception handling, and delivery responsibility.

For enterprise AI services or one-person companies, six asset types are worth accumulating: business task descriptions, well-sourced data and knowledge, portable skills and processes, real failure cases, independent verification methods, and long-term run records.

These assets shift services from "demoing a Bot" to "continuously holding a verifiable responsibility." Industry knowledge, implementation experience, and customer outcomes still cannot be automatically obtained by calling a stronger model.

Grok Build being open source means the runtime system is open, not that Grok model weights are public. The official repository can provide developers framework and interface references; this report verified public statements but did not complete a source-code security audit. [S30–S31, B]

15.3 Individuals: Use Different Roles, Don't Mistake Count for Efficiency

Individuals can divide information organization, development, and project follow-up into roles, but first assess whether omissions and duplicated work decrease. When multiple Bots share state, more roles may also be harder to audit.

Long-term memory should be checkable and correctable; key results should be saved as ordinary documents, code, or business-system records. Individuals controlling their own knowledge and work history can lower future migration costs from supply-relationship or product changes.

15.4 A Minimal Verification Experiment

Select 20–50 real past tasks of the same type, covering simple, difficult, and anomalous cases; reserve some for not modifying processes. Compare complete results of manual methods, single-Agent, and persistent-Agent.

Add human checking and rework to total costs; record delivery correctness, time, anomalies, and actions needing re-authorization. Task sample size is only a practical starting point and doesn't automatically provide statistical significance; high-risk or rare failures need larger scale and dedicated design.

If results improve, expand responsibilities; if improvement only on familiar samples, first study overfitting; if success increases but human time also increases, re-evaluate organizational costs.


16. Key Unknowns and Final Judgment

This study still lacks reliable answers to the following questions:

This report's final judgment [D]: xAI/SpaceXAI's public roadmap already provides an important set of candidates for superintelligence formation mechanisms: base-model growth, AI generating next-generation training materials, Agents forming persistent work containers, development platforms accumulating execution feedback, and industrial resources expanding operable scope. The key unknown is whether these components can produce independently verified R&D acceleration and organizational capability growth.

Therefore, future research should ask separately: whether Grok better solves hard tasks at fixed investment; whether persistent Agents stably lower the total cost of completing work; whether AI research feedback continuously changes next-generation capability; whether infrastructure expansion forms reliable, economical, controllable compute supply.

Only when these questions have continuous data is there basis to raise judgments of Takeoff or system-level superintelligence. Current evidence supports continued tracking; it does not yet support declaring ASI or autonomous R&D takeoff has been realized.


Sources and Verification Notes

A | Independent Measurements, External Research, and Government Materials

No. Source, date, and link Usage scope and limitations
S13 METR: Task-Completion Time Horizons, verified at cutoff date Read methodology notes; 50%/80% boundaries for well-defined tasks like software; no usable independent 4.7 value obtained
S14 Artificial Analysis: Grok 4.7 xhigh, 2026-10-07 snapshot Capability index, API costs, and specifications; ≠ total enterprise benefit
S15 Artificial Analysis: Coding Agent comparison page, 2026-10-07 snapshot Read index v1.5, model combinations, costs, run times, and methodology notes; cites both Grok generations' results, doesn't rely on the page's narrative comparison conclusions
S16 Proximal: FrontierSWE V2, cutoff-date snapshot Read task counts, trial budgets, rewards, and costs; independent of xAI, but the evaluator's full interests not audited
S20 FCC: DA 26-113 acceptance notice, 2026-02-04 2-page full text verified; proves acceptance and comment solicitation only, not final approval
S21 NASA: 7.0 Thermal Control, verified at cutoff date Read heat balance, vacuum heat transfer, radiation, and materials sections; not a SpaceX design assessment
S24 IEA: Key Questions on Energy and AI—Executive Summary, 2026 edition Read industry bottlenecks and power forecasts; global data-center caliber ≠ xAI power use
S25 Gauld et al.: Quantifying AI data center NOx emissions from space, 2026-08-23 draft, 09-15 revision Preprint; verified metadata and abstract, did not reproduce full-text measurements or final regulatory determination
S26 Srikanth et al.: Recursive self-improvement of AI research agents, 2026-09-22 External preprint; verified abstract, did not reproduce; must not serve as xAI internal roadmap evidence
S27 European Commission: Commission investigates Grok and X's recommender systems, 2026-01-26 Official investigation notice; don't use it to declare final illegality or fines
S28 ICO: ICO announces investigation into Grok, 2026-02-03 Read investigation subjects, scope, and procedure; notice stated no final infringement judgment at the time

B | First-Hand Company Disclosures

No. Source, date, and link Usage scope and limitations
S01 xAI: xAI joins SpaceX, 2026-02-02 Official organizational change; don't infer full legal structure from the short notice
S02 Cursor: Cursor is joining SpaceX, 2026-08-14 Official joining statement, April collaboration start
S03 xAI: News, verified at cutoff date Check public release catalog and chronology; don't measure capability by release counts
S04 xAI: Introducing Grok 4.7, 2026-09-21 Read training changes, native framework adaptation, pricing, and releases; comparison tables belong to official disclosure
S05 xAI: Grok 4.7 Model Card, 2026-09 30 pages total; focused verification of identity, training, coding evaluations, and safety sections; not independently reproduced, no full-text audit claimed
S06 xAI: Introducing Grok 4.5, 2026-07-16 Verify Cursor collaboration predates the August joining; don't restate the whole performance table
S07 xAI: Team Bots, 2026-09-28 Read context, memory, engineering, and team cases; headcount, PRs, and timelines are company self-reports
S08 xAI: Designing Grok Bot for a world of persistent agents, 2026-09-03 Read design primitives, roles, events, and interfaces; interface phrasing not proof of isolation architecture
S09 xAI: Workflows in Grok Build,2026-07-23 Read orchestration, budgets, verification, and cross-run improvements; budgets are not measured concurrency
S10 xAI: Memory in Grok Build, 2026-09-16 Read topic files, project scope, and reuse mechanisms; not proof of parameter learning
S11 xAI: Grok Bot products & FAQ, verified 2026-10-07 Current availability, shared computer, user isolation, permissions, and enterprise options; no actual deployment verification
S12 xAI: Using Grok Bot to scale customer support, 2026-09-22 Read the case and gradual authorization; costs not full costs, headcount replacement is the vendor's claim
S17 xAI: New Compute Partnership with Anthropic, 2026-05-06 Colossus access agreement, 220k mixed GPUs, and orbital collaboration interest; not audited inventory or GW construction contracts
S18 xAI: Colossus, verified at cutoff date Historical build-out and multiple device counts; the page has count and timing inconsistencies
S19 xAI: Raises $20B Series E, 2026-01-06 Financing, year-end equivalent compute, and in-vehicle voice product connection; not cutoff-date cash or physical GPU inventory
S21A OpenAI: Our decision on Cursor following its acquisition by SpaceX, 2026-08-28 One-sided contract termination intent and proposed November 12 date; allegations not written as independent rulings
S23 xAI: Frontier Artificial Intelligence Framework, 2026-06-30 public version 9-page full read; risk categories, assessment cycles, and deployment principles; execution not publicly audited
S30 xAI: Grok Build is now open source, 2026-07-15 Runtime framework open, doesn't mean model weights public
S31 GitHub/xAI: grok-build, verified at cutoff date Read public README, license, and sync notes; no source-code audit completed

C | Authoritative Media and Supplementary Verification

No. Source, date, and link Usage scope and limitations
S22 Reuters: SpaceX aims to launch orbital AI computing tests, 2026-06-09 Reporting summary/obtainable paragraphs; future test plans, not realized results
S29 WSJ: Can AI Run in Space? Google Is About to Find Out, early October 2026 Reporting summary; only for industry-having-entered-prototype-testing background, not attributed to xAI already running GW compute

Research Coverage and Methodological Limits

This study searched official products, model cards, development frameworks, independent evaluations, R&D research, energy, space thermal control, and regulatory materials. It followed a "material collection → evidence map → competing explanations → causal synthesis → falsification → report" process, with the report structure organized by formation mechanisms after verification.

Obtainable content includes full web pages, selected PDF paragraphs, complete policies and government notices, and also clearly marked paper abstracts and media summaries. Abstract verification was not written as full-text study; vendor cases were not written as independent reproduction. Public benchmarks may use different model tiers, frameworks, and task versions; this report only performs generational calculations within the established caliber of the same comparison page.

The model–system–infrastructure framework, economic formulas, heat-dissipation magnitudes, scenarios, and monitoring thresholds in the report are all D-grade analysis, not vendor roadmaps or experimental results. It did not access the lab's internal systems and does not claim knowledge of undisclosed R&D progress.

Quarterly update order: first save new materials' dates and versions; update specific indicators in the tables; record conflicts and unknowns; then adjust scenarios and core judgments. Keep old snapshots; don't overwrite old trends with new baselines.


End of report | GPT V1.0 | October 7, 2026