As Muse and Astra usher in the era of “digital workforce,” Elon Musk sets his sights on “AI computing power supply”! SpaceX (SPCX.US) turns AI competitors into major clients.

According to news from Zhihu Finance APP, SpaceX (SPCX.US), founded and led by Elon Musk, and known as a leading company in both "AI + Space Exploration" with a market cap approaching $2 trillion, appears to be shifting its "AI grand ambitions" from competing for AI large model supremacy to broadly selling or leasing its colossal AI computing power resources to the global AI industry. In a report released by The Information on October 1, insiders revealed that the originally in-house AI supercomputing clusters used by Grok have transformed into a "Neocloud"-type AI compute rental or sales business that serves the demands of major AI large model developers and cloud computation clients such as Anthropic and Google.

It is understood that SpaceX is planning to launch about 420,000 Nvidia GPUs in November, expanding its Memphis Minihard campus and preparing capacity for a new agreement worth about $1.1 billion per month, starting in December. In the summer, they also discussed leasing computing power with Microsoft, but the outcome is not yet clear. As competition among models continues to drive demand, SpaceX is trying to turn its high-efficiency AI infrastructure build-out, robust power resources built around AI systems, and AI cluster operation capabilities into a strong revenue curve. This is also why the market has started to value SpaceX’s “AI computing power valuation” more, pushing the company's share price up nearly 40% since the lows in early August, and bringing its market cap close to $2 trillion.

Google, a subsidiary of Alphabet, also took part, signaling a very positive sign for the AI computing power supply chain: Per disclosed agreements, it will receive approximately 110,000 Nvidia GPUs, matching CPUs, DRAM memory/NAND data center storage systems, and other massive resources, paying $920 million monthly from October 2026 to June 2029. Even this tech giant with its proprietary TPUs and vast cloud infrastructure needs to purchase external compute capacity, highlighting that the battle for computing power is also a race for available capacity and go-live dates.

Micron, one of the world’s top three memory chip manufacturers, issued strong results and outlooks that further crystalize this unprecedented AI infrastructure investment boom in the profit statements of compute power suppliers: In Q4 FY2026, Micron’s revenue reached $54.229 billion, up roughly 379% year-on-year, with adjusted EPS at $33.42; the next quarter guidance is for $61.5 billion +/- $1.5 billion. Data center SSDs quarterly income is approaching $10 billion, more than ten times the previous year, and most HBM supply for 2027 has already been signed with significant price increases. Management expects memory supply and demand from 2027—2028 will become even tighter than the record-tight current market, proving that the demand for AI computing is expanding the total commercial value of the entire AI supply system, including memory layers.

The pivotal transformation at SpaceX is converting previously underutilized resources into sellable services. The Information's latest report points out that Colossus's utilization rate was once below 40% last year and remained low in the spring; Musk initially opposed leasing capacity, later accepting demands from Anthropic, including fully activating H100 resources that were of relatively low efficiency for Grok’s own training. The Information notes the company has also solved multi-tenant security isolation in the original system. Between chip ownership and cloud service profits, there are engineering thresholds such as scheduling, software, security isolation, and customer access; only by overcoming these can partial idle capacity be converted into the urgently needed industry supply.

For faster delivery, the company has stockpiled millions of dollars’ worth of gas turbines, brought in ABB robotics to organize wiring, and employs modular construction in parallel—so called “electricity locking, compressing construction cycles, and patching reliability.” It is also accelerating its 500MW data center with Saudi Humain. However, drainage and structural issues with Colossus II could lead to at least two to three weeks of delays, prompting SpaceX to rotate engineers, enhance project management, and boost reliability.

All these details collectively illustrate—what really determines the commercial scale of AI computing resources is the full supply chain capability—from power supply, construction, to stable operations. This is exactly SpaceX’s AI compute aggregation advantage; clients are buying always-on compute services, so delivery must cover the entire chain, from electricity to cluster stability. Anthropic and Google, both long-time competitors of Musk’s xAI (now acquired by SpaceX) in the AI large model/agent space, are now the largest customers of SpaceX’s computing resources. This enables SpaceX to capture strong and growing AI compute demand from multiple model ecosystems and agent systems in the AI super agent era, boosting utilization rates and diversifying revenues.

The Involvement of AI Agents in Enterprise Operations: How Sequences of Commands Activate Entire AI Data Centers

The latest IPO prospectus from global AI large model/application leader Anthropic shows, from the computing resource demand side, why SpaceX is undergoing this transformation, with the most industrially meaningful part being the expansion of actual compute cost. Previous media leaks of Anthropic’s confidential IPO paperwork revealed that Anthropic’s 2025 revenue is close to $4.6 billion, with operating losses at $8.06 billion, and $7.33 billion spent on computing and infrastructure; about triple the prior year, constituting roughly 58% of total $12.65 billion operating expenses. This newly released IPO data indicates that frontier model commercialization is expanding both revenue and resource consumption, with AI agent/model training, inference, and infrastructure investment far outpacing profit realization.

More long-term signals come from Anthropic’s at least $518 billion infrastructure commitments over the next decade, about 80% of which are non-cancelable or require payment regardless of usage; arrangements with SpaceX/xAI of up to $84.5 billion by 2029 can mostly be cancelled with 90 days’ notice. Note these are future contractual obligations, which should be understood separately from current outlays and recognized revenues. Yet all these developments point to the same trend—the world’s top AI large model labs are making long-term commitments for future capacity, and SpaceX is trying to turn build-speed into a stronger supplier position.

According to sources cited by the media, Anthropic has signed a deal worth up to $84.5 billion to use SpaceX’s AI compute resources through 2029. This means Anthropic’s payments to SpaceX for computing power are nearly double previous estimates, emphasizing that the faster the commercialization of AI large models/agent workflow, the higher the strategic value of compute clusters that can be delivered reliably and on time.

Recent technical advances led by Muse, Astra, and Anthropic Claude have laid the critical foundation for the large-scale commercialization and explosive demand for AI compute power across industries—especially as the application scope of agents expands, driving synchronous demand for AI GPUs/TPUs, high-performance CPUs, high-performance HBM/DRAM/NAND data center memory chips, and high-speed optical interconnects—the core infrastructure of AI.

From an AI inference system architecture perspective, GPUs and dedicated accelerators handle model computation, CPUs translate inference results into practical action, and memory/storage retains and recalls task state. Long-context and multi-turn calls increase pre-fill calculations and KV (key-value) cache demand; browsers, code sandboxes, retrieval, and task orchestration boost server CPU load; files, databases, persistent memories, and tiered caches push demand on server DRAM, enterprise SSDs, and high-speed networks. Nvidia's technical documentation now describes AI agent inference as an immense systems engineering challenge spanning GPU HBM, CPU DRAM, local NVMe/remote storage, and the internal high-speed optical interconnects crucial for data movement.

In Anthropic’s management view, memory chips in AI data center server clusters, as well as AI GPUs, remain the clearest bottlenecks in the AI computing industry supply chain. Research firm TrendForce expects server DRAM contract prices to rise about 270% cumulatively by 2026, enterprise SSD prices by 235%; HBM contract prices could jump another 70%—140% by 2027, continuing their doubling trajectory. These numbers reflect both rising demand for AI computing power and storage chip price surges. TrendForce also estimates that in 2027, shipments of NVL72 racks, including Blackwell and Vera Rubin platforms, will increase by over 50% from the previous year; their market research visual shows system value rising from $226 billion in 2026 to $711 billion in 2027, a 214% YoY increase.

The core shift brought by Muse and Astra is to make "digital labor" the new primary consumer of compute power. Meta’s Muse is driven by Muse Spark, with a dedicated cloud VM that continues working after a user leaves the app, overseen by an independent Sentinel agent reviewing all external operations; OpenAI’s GPT-6 Astra powers Dots, where each agent has its own cloud-based computer, browser, and tools, capable of ongoing work and spawning sub-agents. Thus, a single user command triggers a chain reaction of planning, retrieval, execution, validation, and re-inference.

The overall mechanism behind surging AI compute demand can be summed up as—Total inference workload ≈ Number of active users × Tasks per user × Model calls per task × Compute per call. Increased penetration acts on the first term, workflow automation and multi-agent collaboration on the middle terms. Prior research by Anthropic noted agents use about 4x as many tokens as standard chat, while multi-agent systems use about 15x; while these empirical numbers can’t be directly projected onto Muse or Astra, they do illustrate the leap from “answering questions” to “efficiently accomplishing human-complex work.”

The deeper driver for growth is the declining cost of each successful task, expanding the economically viable scope of AI’s work. Astra, for instance, improves both capabilities and token efficiency in some public benchmarks, meaning demand growth doesn’t require each task to further increase resource drain—so long as added task volume exceeds resource savings per task, total compute demand keeps expanding. Furthermore, competition between open and closed large models, through quality upgrades, lowered barriers to use, and scenario expansion, jointly enlarges the market for AI computing resources.

The "Fulfillment Kill Line" of the AI Frenzy: How SpaceX Opens Up a $3 Trillion Valuation Imagination

The rapid spread of Muse and Astra’s AI agents is shifting data center resource structures. GPUs now do most model computation; CPUs handle browsers, code sandboxes, tools, data processing, and task orchestration; long context and high concurrency spur higher KV cache demand, driving HBM capacity and bandwidth upgrades; DDR supports VMs, tool processes, and some cache offloads; enterprise SSDs are for files, long-term memory, and tiered context storage. Pre-fill/decoding separation across nodes means KV caches need to move over high-speed networks, boosting demand for networking equipment and optical interconnects. Effective compute power depends on the scalability of the full system.

This also explains why the investment theme is spreading toward CPUs, storage, and intra-data center high-speed optical connectivity. On September 21, the Muse download frenzy helped push AMD up 10%, Intel 12%, and Arm 17%. Korea’s official September data (the country is home to SK Hynix and Samsung, global storage giants) underscored the trend: semiconductor exports at $60.3 billion, up 262.8% YoY; computer exports at $7 billion, up 435.3%; semiconductor equipment imports at $3.44 billion, up 50.8%. Export value was boosted by both volume and price, while increased equipment imports signal compute demand is feeding through to wider semiconductor capacity expansion.

Wall Street analysts, spotting this trend, are proposing a full-industry judgment: A global AI buildout wave is creating an “AI computing fulfillment kill line”—the commercial ceiling now depends on how much successful task output enterprises can deliver, at what cost, latency, and reliability. SpaceX’s early power securing, cluster build-out, client acquisition, and reliability enhancement are all about claiming a position at this supply boundary. For capital markets, re-rating metrics have become more concrete—namely available power, on-schedule capacity, paid utilization rates, effective task throughput per MW, and ultimately, cash flow generation.

According to information, top Wall Street firm TD Cowen initiated coverage of SpaceX at the end of September with a "Buy" rating and a target price of $200, implying about a 35.1% increase on the October 1 close of $148.07; their core logic is that ground-based AI compute rentals will become the primary growth engine in the near term, projecting $66 billion in related revenues in 2027, about 58% of total revenue, and making up over half from Q1 that year. Among other Wall Street giants: Goldman Sachs rates it "Buy" with a 12-month price target of $220, Morgan Stanley rates it "Overweight" with a target as high as $300, UBS rates it "Buy" at $210, and Bank of America rates it "Buy" with a $235 target.

S&P Global aggregated 37 Wall Street analyst submissions as of September 30, with an overall rating of "Buy," the most optimistic judgement, and a 12-month average price target of $226.04, representing a 52.7% upside from $148.07; given static shares outstanding of about 13.57 billion, this implies a market value of about $3.07 trillion, higher than the current $2.01 trillion. Analysts' expectations for this optimistic valuation scenario hinge on SpaceX, under Musk’s leadership, pivoting ever more strongly toward converting power acquisition, build speed, and cluster operation capabilities into continual revenue streams from servicing multiple AI clients.