GPT-6 Astra Resets the Cost Curve

GPT-6 Astra Resets the Cost Curve as Frontier AI Debates Its Own Speed Limit

OpenAI released GPT-6 Astra last week in a soft launch, with wider availability to follow. The reported benchmarks were strong, scoring 90% on MathArena, 97.6% on FrontierMath Tier 4 and 95% on a robot-arm task against 40% for a rival frontier model. The argument ran that if even half the implied timeline acceleration holds, capex into AI infrastructure, robotics and compute could shift into the next two quarters rather than 2028. The prevailing read was more measured. Astra's real edge is cost: it solves the same hard problems with far fewer tokens, just as enterprise finance teams cap AI budgets, reinforced by compute contracts locked in early at roughly 20% of today's going rate that could hold a structural cost advantage through 2027.

That efficiency gain matters more than usual because compute scarcity has refused to ease. Second-hand prices for older GPUs rose rather than fell over the summer, defying the depreciation schedules baked into hyperscaler and neocloud balance sheets. Microsoft moved to more than triple its compute capacity after turning away business, and GPU renewal pricing pushed through at double-digit increases. If near-insatiable demand keeps residual values firm, the useful-life assumptions embedded in installed-compute models may prove too conservative, flattering returns on hardware already deployed. A model that delivers equivalent output with fewer tokens therefore compounds the advantage for any operator sitting on scarce, appreciating silicon.

The efficiency carried a transparency cost: much of Astra's processing goes unreported, making the model harder to audit, and the AGI framing drew scepticism, with the counter that this is a genuine improvement rather than the arrival of silicon intelligence.

The broader backdrop was a live debate on whether frontier development itself should slow. Anthropic's Dario Amodei published an essay calling on the industry to moderate the pace, pointing to the emergence of recursive self-improvement over the summer, where models begin building better models with diminishing human involvement, and warning that within six to twelve months a swarm of misaligned agents could cause hundreds of billions of dollars of damage. OpenAI's Sam Altman had separately floated a similar go-slow, so the direction of travel was broadening. The sceptical read was that the doomerism partly served commercial ends, building anticipation ahead of Anthropic's IPO and nudging Washington toward curbs on open-weight competition. For infrastructure investors, the risk is that a collective go-slow moderates the pace of capex, which would weigh on the pick-and-shovel layer of the AI supply chain. The counterweight is that the bull case rests on parabolic token consumption as AI permeates the economy, which can persist even if the bleeding edge advances more slowly. A cheaper, stronger OpenAI release landing right before Anthropic's listing, with Meta's latest model keeping the token-price war live, ensured the competitive pressure remained firmly intact.

What Drove The Collective Ideas Last Week

Collective Conviction Ideas: -1% WoW

The Conviction ideas declined around 1% in aggregate last week, with direction split by duration and catalyst. Data-centre silicon and frontier-market digitalisation led the upside as company-specific developments provided support. On the downside, long-duration power infrastructure bore the heaviest losses, the digital-asset sleeve softened alongside a weaker crypto tape, and European cyclicals handed back recent gains. The pattern was consistent: short-duration, cash-generative positions with near-term catalysts held up, while longer-duration exposures lagged.

Find out more about Collective ideas on CurationAI

Best Content Shared This Week

🎧 Jeff Currie: The Great Rotation Into Hard Assets

Why Listen? Former Goldman commodities chief Jeff Currie argues we are only in the “second or third inning” of a generational commodity supercycle, driven by sovereign debt, currency debasement and chronic underinvestment in physical assets. From his $10,000 gold call and “Munificent Seven” energy majors to uranium, agriculture and critical minerals, the discussion lays out why capital could continue rotating away from financial assets and towards the physical economy for years to come. Listen here.

📚 Nebius Moves Inside the Palantir Stack

Why Read? Palantir has named Nebius its preferred sovereign AI infrastructure partner, giving the neocloud direct access to Palantir's enterprise ecosystem and bringing its compute inside the same perimeter as customers' data and models. The piece explores why combining Palantir's software and distribution with Nebius's AI-native infrastructure could strengthen the sovereign-AI thesis and open a significant new route to enterprise growth for Nebius. Read here.

Stock Of The Week

Building Europe's Minerals Platform

SigmaRoc is a European lime and limestone group supplying essential materials into industrial, environmental and construction markets. The first half reinforced the underlying strategy: build a portfolio of hard-to-replicate mineral assets, improve their operations and use the resulting cash generation to consolidate a fragmented European market. The group now operates around 110 sites across 14 countries, backed by roughly 2.7bn tonnes of mineral reserves.

That consolidation story moved forward again last week with the acquisition of AB Dolomitas, announced alongside SigmaRoc's H1 results. The deal adds another strategic minerals business to the platform and fits directly with management's stated model of acquiring scarce, asset-backed businesses where permitting barriers and long customer relationships provide protection from new competition.

The attraction is the combination of organic improvement and M&A. SigmaRoc is targeting 3–5% annual organic growth, EBITDA margins above 24% and ROIC above 15%, while retaining a significant acquisition runway across Europe. With the first-half performance showing the existing portfolio delivering as another asset is added, the investment case is increasingly about whether SigmaRoc can keep compounding through the same buy, improve and integrate playbook.

Explore SRC LN

New or Updated Showcases

Pan Global Resources

• Pan Global Resources is a copper-focused exploration company advancing the Escacena and Cármenes projects in Spain, giving investors exposure to potential new European supply of a metal increasingly critical to electrification, grids and AI infrastructure. Its flagship Escacena project hosts the La Romana copper-tin-silver discovery alongside a growing pipeline of nearby targets, positioning the company to benefit if continued drilling converts exploration success into a larger, strategically located resource. Read more.