Elon Musk signals continued large-scale GPU ordering as Tesla readies an in-car inference processor, underscoring the split between edge autonomy and data-centre training and the pressure on margins as rival carmakers standardise on platforms.
Merifund Capital Management is flagging a clear message from Tesla’s leadership: Nvidia remains core to the computing stack behind autonomous driving, even as Tesla advances an in-house processor for vehicles and robots. That stance carries weight as industry forecasts point to roughly $1 trillion of additional AI chip revenue over the next two years.
Tesla’s chief executive uses the X platform this week to confirm that Tesla and SpaceX AI continue ordering Nvidia semiconductors at scale, signalling that proprietary development is not a near-term substitute for external capacity. The disclosure keeps attention on Tesla’s planned chip-manufacturing initiative, scheduled to begin in the coming days, as the company sets out a longer-term route to more internal control.
Anthony Saunders, Director of Private Equity at Merifund Capital Management Pte. Ltd., characterises the approach as “a dual-track build that protects delivery while the silicon matures”, with data-centre GPUs doing the heavy lifting for training and the new AI5 chip focusing on edge inference. Saunders presents the distinction as central to understanding the strategy: training consumes vast compute for repeated model iteration, while inference prioritises low latency and power efficiency inside a vehicle.
On Tesla’s stated targets, AI5 delivers around 40x to 50x more compute and about 9x more memory than the current AI4 generation, while drawing roughly 150W at rated power versus around 700W for a high-end data-centre accelerator. Tesla positions the chip for robotaxis and Optimus, where a dedicated neural design can process sensor inputs locally without relying on a remote data centre for split-second decisions.
The economics of autonomy depend on what happens before a model ever reaches the car. Tesla’s fleet generates more than 10 terabytes of driving and sensor data each day, and training requires centralised clusters that can ingest and learn from huge datasets. In its current reported configuration, Tesla deploys 5,760 Nvidia A100 GPUs to train on around 1.5 petabytes of driving data in the active training corpus, and Merifund Capital Management treats that capability as a critical bottleneck. Saunders describes continued procurement as “the insurance premium that keeps the programme moving when timelines shift”.
Manufacturing remains the longer pole. Tesla’s chip facility targets 100 billion to 200 billion chips a year at full run-rate, with a projected budget of roughly $22.1 billion to $29 billion over the construction and tooling phase. A reported $18.3 billion manufacturing agreement with Samsung, alongside additional foundry routes, supports resilience, yet volume output for automotive-grade silicon still points to a multi-year ramp, keeping Nvidia’s role operational across the next several reporting periods.
Nvidia’s platform strategy sharpens competitive pressure across the sector. At its flagship conference this week, Nvidia positions DRIVE Hyperion as a production-ready route to higher levels of autonomy, with BYD, Geely, Nissan and Isuzu joining Mercedes-Benz, JLR and Volvo Cars in adopting the stack. Uber outlines an ambition to deploy DRIVE Hyperion-powered robotaxis across 28 cities over the next few years. Saunders views the shift as “the commoditisation risk investors need to price, because platform availability raises the baseline for rivals”.
The valuation debate now hinges on timing and cash. Over the latest reported quarter, Tesla’s operating margin sits near 8.8%, while analysts model AI infrastructure spending that can weigh on profitability as capacity expands. Current projections put capital expenditure above $22.1 billion over the next fiscal period, excluding the chip facility, and estimates for the semiconductor build-out reach about $33.2 billion, while some forecasts also anticipate free cash flow around -$4.5 billion over the same horizon. The latest bank note values autonomous driving at about $298.9 a share, implying roughly $1.3 trillion of equity value using a fully diluted share count, even as Tesla’s market capitalisation hovers around $1.2 trillion on current prices and its earnings multiple sits above 200. Saunders calls the spread “a market test of execution, repeated every quarter”.
For Merifund Capital Management, the immediate takeaway is that Tesla’s chip narrative is not a clean pivot away from Nvidia, but an attempt to blend proprietary edge hardware with proven training capacity. That balance keeps supplier dependency visible, keeps capital intensity in focus, and leaves investors watching a short list of live signals, including production ramp cadence, software adoption and the tempo of GPU orders.
Founded in 2010, the firm (UEN: 201024554E) is a Singapore-headquartered hedge fund manager with strategies spanning traditional long-only mandates, long/short equity, global macro, event-driven and systematic trading. The firm uses derivatives to pursue opportunities while prioritising liquidity, capital preservation and disciplined risk controls, and it incorporates ESG considerations in line with global sustainability standards. It serves accredited investors, family offices, foundations and endowments, and it is expanding access for retail investors. Further insights are published at https://merifund.com/insights. Media enquiries can be directed to Tao Yang at media@merifund.com, with additional information available at https://merifund.com.