Over the past three years, I have been watching the center of gravity in LLM competition shift. In 2023, the bottleneck was compute chips—having high-end GPUs was enough to seize the early lead. From 2024 to 2025, everyone’s attention turned to the scale and architecture of foundation model weights. By 2026, the true watershed has fallen on the behavioral trajectories left by human interaction in real-world work scenarios.
This data is fundamentally different from static web corpora scraped from the public internet, and completely unlike synthetic text churned out in bulk by models. It carries distinct physical properties: web crawlers cannot find it on the public internet, algorithms cannot conjure it out of thin air, and it can only precipitate out of the actual scenes where real work takes place.
On August 14, 2026, SpaceX completed its acquisition of Anysphere, the parent company of Cursor. The entire $60 billion transaction was paid in stock, with Anysphere shares converted into 389,289,254 shares of SpaceX Class A stock—I previously analyzed the option structure and IPO motivations behind this deal. The groundwork for this deal was laid early on. In February 2026, SpaceX merged with xAI to form SpaceXAI. When the two sides partnered in April, an exit mechanism was already locked in: SpaceX would either acquire the company outright for $60 billion, or, if the partnership fell apart, pay a breakup fee of $1.5 billion in cash plus $8.5 billion worth of compute resources.
The chips in both hands complemented each other perfectly.
SpaceXAI built the Colossus supercomputer cluster in Memphis, Tennessee, equipped with roughly 200,000 Nvidia GPUs and plans to expand to 1 million GPUs—the context surrounding xAI’s compute commercialization and leasing Colossus to Anthropic/Cursor has been discussed previously. Colossus has an extraordinarily deep compute pool, but lacked the continuous interaction records of developers during daily coding and debugging.
Cursor’s situation was the exact opposite. As a leading AI coding tool, Cursor served more than 50,000 enterprise customers, covering 64% of the Fortune 500, with annualized revenue surging from over $1 billion at the end of 2025 to around $4 billion by June 2026. Yet Cursor researcher Federico Cassano once revealed that their team owned only tens of thousands of GPUs, and their inference infrastructure had long been hosted on third-party Fireworks, relying on 3 to 4 globally distributed clusters to run distributed rollouts and inference services.
Sarah Wang discussed the special nature of this data in Sarah Wang’s analysis on X. Cursor not only holds finished code, but more importantly captures the full trajectory of how top-tier developers conceptualize, experiment, encounter errors, and course-correct in their actual work—see the previous analysis of Cursor’s cognitive lead.
aiweekly.co’s report framed this union as a flywheel: session data accumulated by Cursor is fed into model training; the newly trained foundation model is plugged back into the editor, attracting more developers to write code and generating even more granular data. Fireworks’ official blog similarly pointed out that mastering this closed-loop cycle is becoming a decisive barrier in AI systems competition.
This complementarity was quickly reflected in product evolution. In July 2026, Grok 4.5, a 1.5-trillion-parameter model jointly developed by xAI and Cursor, was launched. According to mightybot.ai, the training process consumed trillions of tokens of Cursor developer interaction data. Grok 4.6, released on August 12, continued using this 1.5-trillion-parameter foundation. According to reporting from develeap.com, the performance gains in 4.6 primarily stemmed from the post-training stage, including longer continued pre-training, regenerated fine-tuning data, and agentic reinforcement learning, with benchmark comparisons documented in the vktr.com benchmark comparison.
However, there is an attribution question that the entire industry is watching: public information makes it difficult to prove just how much of Grok 4.5 and 4.6’s capability gains actually came from Cursor’s interaction data. Without rigorous controlled experiments, the performance improvements could just as easily be explained by stacking compute on the Colossus cluster for extended training runs.
Even so, SpaceX’s move sent a clear signal. The highest-tier developer behavioral data lives inside mature, externally operated independent products. If a buyer wants access to this data portal, the most direct solution is to buy the entire product along with its flywheel in one fell swoop.
If Cursor represents a high-speed, live online flywheel, another deal demonstrates that the digital inventory of a shuttered company carries just as clear an asset value. According to a Reuters report on August 17, 2026, Google bid $10 million in Spirit Airlines’ bankruptcy restructuring proceedings to win the airline’s digital asset package. Google didn’t want airplanes, nor did it want flight routes. It wanted something else: the massive operational records accumulated by a mid-sized airline over years of daily business.
Opening up this asset inventory reveals almost every slice of collaboration in a modern enterprise: approximately 100 million emails, 500 million Teams instant messages, 17.08 million OneDrive documents, 20.58 million SharePoint collaborative files, and roughly 30 million lines of internal code across 516 repositories. The inventory also includes daily operation logs, historical fare change records, HR files, ticket routing histories, and diverse business workflow data. Google stated publicly that this data would be used directly for product R&D and AI model training. Interestingly, AI startup Mercor also participated in the bidding, submitting a $7.5 million offer to become the backup bidder, showing that AI teams across different domains are hunting for workflow data from real enterprise environments.
The transaction is currently awaiting final ruling at a bankruptcy court hearing on August 19, 2026. The restructuring agreement is tightly drafted: before the asset package is delivered to the buyer, an independent third party must complete de-identification, and documents protected by attorney-client privilege are excluded from the sale.
Even when a physical company enters bankruptcy restructuring, the digital traces left behind by its day-to-day operations can still fetch tens of millions of dollars at a court auction. Massive, authentic, and intricate enterprise collaboration records are locked inside a defunct entity, and bankruptcy proceedings have become a conduit for buyers to acquire these dormant working memories in one comprehensive package.
While bankruptcy auctions represent a one-off mining of historical inventory, actively operating software vendors are building continuous data pipelines through another approach. On August 17, 2026, enterprise collaboration software provider Atlassian enacted a new data contribution policy, see the Atlassian policy overview. This policy explicitly collects eligible metadata and in-app data from Jira, Confluence, Jira Service Management, platform apps, and select Teamwork Graph connectors to improve product features and AI experiences for all customers.
A closer look at this policy’s design shows that Atlassian directly tied control over data contribution to the enterprise’s software subscription tier, with specific terms outlined in the tier settings documentation. In the Free and Standard basic tiers, metadata contribution is mandatory and cannot be disabled by customers, while in-app data collection is enabled by default but can be manually turned off. Moving up to the Premium tier, metadata contribution still cannot be disabled, but in-app data collection becomes opt-in (disabled by default). Only enterprise customers purchasing the top-tier Enterprise plan have full autonomy to opt out of metadata sharing entirely.
To support the rollout of this policy, Atlassian updated its Customer Agreement, AI Terms of Service, Data Processing Addendum, and Privacy Policy on the same day. The official statement specifically explained the processing mechanism: when collecting metadata and in-app data, the system extracts common phrases, keywords, and business themes, protecting privacy through de-identification and aggregation while filtering out low-frequency features that could directly identify a specific enterprise.
Tens of thousands of enterprise employees collaborate and drive projects forward every day inside Jira and Confluence. By adjusting its terms of service, the vendor has seamlessly turned its customers’ routine software usage into the input stream needed for continuous model iteration.
When observing DeepSeek’s product cadence previously, I always had a question: for coding agents, they mainly open-sourced model weights and provided API endpoints, leaving the interaction interface and execution environment to external third-party tools—I previously analyzed the role of the harness as a behavioral data portal.
This lightweight model moved extremely fast during early expansion, but it also created an obvious data blind spot. At the API backend, the model provider receives only isolated input and output requests, completely blind to how developers break down tasks step-by-step in local terminals, which step threw an error, or how many rounds of edits were made before the code finally ran successfully.
Without this layer of fine-grained micro-interaction traces, optimizing the model for complex multi-step tasks is like looking through frosted glass. According to LangChain test data, the same foundation model can see a 10 to 20 percentage point gap in complex task execution success rates depending on whether it is paired with a dedicated harness versus lacking targeted configuration. DeepSeek recently began recruiting for harness-related roles in the market, clearly aiming to fill in this crucial front-end antenna. If there is no suitable data portal on the market or it cannot be secured through commercial agreements, a team must step in to build the full front-end interaction surface and execution environment itself, accumulating proprietary behavioral feedback from real user interactions.
Looking at these four cases together, the four paths to acquiring organic work data map onto a clear coordinate system across two dimensions: data source and acquisition frequency:
Google’s bid for Spirit Airlines’ data represents a one-time purchase of external historical inventory; Atlassian’s terms update represents continuous accumulation of external incremental data; DeepSeek building its own harness relies on its own ecosystem to nurture a continuous internal closed loop; and SpaceX’s acquisition of Cursor uses capital to bring an established external tool entirely under its wing, directly taking over a data flywheel that is already running at full throttle.
Connecting these cases makes the substantive shift in the industry’s competitive focus clear. The competitive moat for large models is shifting away from model parameters and compute scale toward human behavioral data generated in real work.
This type of data possesses strong physical exclusivity. It cannot be scraped from the public internet, nor can it be generated out of thin air detached from specific business workflows; it can only be produced during the unfolding of real work.
These four acquisition paths are not abstract theoretical categories. Teams are choosing the path of least resistance based on where the data physically resides: launch a capital M&A when data precipitates inside an external independent product; join a judicial auction when it is sealed inside a bankrupt organization; adjust terms of service when it flows through daily enterprise collaboration; and build it yourself when no ready-made surface exists in the market.
For builders of AI systems, planning data assets can no longer stop at simply enabling backend logging. Acquiring organic work data is evolving into an integrated decision that spans capital operations, contract design, compliance boundaries, and product engineering.
Each of the four paths carries clear trade-offs: full acquisition quickly absorbs a mature data closed-loop, but demands massive valuation premiums and organizational integration costs; acquiring external data through terms of service has negligible marginal cost, but readily triggers enterprise customer concerns over data sovereignty and privacy leakage; bidding for historical data in bankruptcy proceedings rapidly supplements cold-start samples, but it is a one-time static snapshot lacking subsequent dynamic updates; and building a harness in-house crafts a proprietary channel tailored to model optimization, but requires weathering long cold-start cycles and the market risks of building a standalone product.
When executing these strategies, builders also need to take a sober look at three practical boundaries that remain insufficiently verified. The first boundary is the true attribution of model capability leaps. In the benchmark gains of Grok 4.5 and Grok 4.6, public information cannot separate how much was contributed by Cursor developer interaction data versus extended training on Colossus’s massive compute cluster. Objectively, empirical evidence with rigorous controls is still lacking on how much data flywheels actually drive large model performance jumps. The second boundary is the retention rate of effective signal after data de-identification. Spirit Airlines’ asset package must undergo strict third-party sanitization before delivery. Erasing privacy and legal risks guarantees compliant delivery, but it may also shatter the coherent causal chains in the original business workflows. Just how much high-quality signal this data will ultimately yield for model training remains an open question. The third boundary is compliance friction during actual enforcement of terms of service. While Atlassian designed tiered rules on paper, whether enterprise customers can exercise opt-out rights without friction and whether aggregated de-identification schemes can withstand rigorous security audits in practice will require longer real-world validation.
As algorithmic architectures converge and compute increasingly becomes a commoditized infrastructure, the competitive barrier for models is inevitably gravitating toward real-world work behavioral data. The sole source of this data is the workplace itself. For builders of AI systems, whoever embeds their products and agents deepest into actual business workflows will command the most critical means of production in the next phase of the race.