Anthropic drops plan to acquire AI startup Decart, Bloomberg reports

Anthropic has abandoned its pursuit of AI startup Decart AI after conducting due diligence on a potential transaction reportedly valued at about $6 billion. Decart develops software intended to improve chip efficiency and lower AI-model training and inference costs, an area relevant to Anthropic as it increases computing-capacity spending ahead of a potential IPO. The companies may still explore collaboration opportunities, but no acquisition agreement was finalized.

Analysis

The abandoned transaction is more informative about private AI infrastructure valuation than public-company earnings. A roughly $6B discussion around software that lowers inference/training cost implies strategic buyers are assigning substantial option value to compute-efficiency IP, but the failure to close suggests either valuation, technical diligence, or integration rights did not clear the hurdle. That is a modest negative read-through for late-stage AI infrastructure private multiples over the next 1-3 months, not a direct signal for GS.

For public AI suppliers, efficiency software has two opposing effects: it can reduce compute required per model run, but lower unit costs generally expand usage and accelerate deployment. Near term, the latter remains more important for NVDA and hyperscaler capex; a single unconsummated private deal does not change demand visibility. Over 6-18 months, however, broad adoption of model-optimization tools would shift value from raw accelerator capacity toward cloud platforms and software layers that can capture the savings, favoring MSFT, GOOGL and AMZN relative to a pure hardware-duration trade.

The contrarian interpretation is that Anthropic may prefer partnerships or internal development because efficiency tooling is not sufficiently differentiated to justify an acquisition premium. If that proves true, the market may be overestimating the scarcity value of private AI-infrastructure assets. The key falsifier is evidence of a competing strategic buyer at a comparable valuation or independently disclosed performance gains that materially lower cost per token/model-training cycle.

Key Decisions for Investors

• No directional position in GS: the item has no identifiable impact on Goldman earnings, capital markets activity, or valuation; treat any GS price reaction as unrelated noise.
• Maintain NVDA exposure rather than reducing it on compute-efficiency headlines; reassess only if major cloud customers guide AI capex lower or disclose falling accelerator intensity. A 1-3 month catalyst is hyperscaler earnings/capex commentary.
• Watch for Decart fundraising, licensing, or a competing acquisition offer as a private-AI valuation indicator. A deal at or above the previously discussed valuation would support software/infrastructure scarcity and favor selective AI-platform exposure.
• For a 6-18 month relative-value theme, consider a staged long MSFT/GOOGL basket versus a hedged semiconductor-equipment or accelerator-overweight basket only after verifiable evidence that inference cost reductions are translating into higher cloud workloads rather than lower hardware purchases.