The Cost of Staying First | Part III: The Race Nobody Wants to Stop

Listen to this podcast

Narrated by Charlotte · The Noble House

The most consequential promise in AI is that a better system will help build the system that succeeds it. If that cycle accelerates, leadership becomes more than a product advantage. It becomes an advantage in producing the next discovery.

OpenAI's September account of research acceleration describes agents being used throughout its research work. Dario Amodei's pacing proposal identifies AI-assisted development of subsequent generations as a reason to change how the frontier advances. Both place research itself inside the transformation they are discussing.[1][8]

That is the setting for the race toward artificial general intelligence and, beyond it, superhuman general intelligence. In this report, AGI means broad competence across intellectual tasks; SGI means general capability exceeding human performance. The strategic issue is what such systems enable their owners to discover, build and deploy. Intelligence becomes power through those results.

Our forecast for 2026–2028 is an intensifying contest across the entire chain: frontier models, open-weight alternatives, devices, chips, energy and industrial production. Greater capability will spread across more applications even as control of some foundational resources remains concentrated. Members should look for opportunities where that spread changes what they can accomplish, what they must pay, and which dependencies they can escape.

Report-wide evidence · 2026-09-13

Compass Predictive Analytics

Signal gauge

76%

Evidence Reliability

12 Of 12 Validated Assertions Have Complete Exact Span And Ownership Lineage. · Positive

tracked

Quantifies the conservative evidence floor after exact-span and independent-owner checks.

11 evidence references

Signal gauge

Evidence Freshness

Evidence Freshness Is 2 For The Report's Validated Factual Foundations. · Negative

tracked

Separates current evidence from aging context using a declared decay window.

7 evidence references

Open models are changing the terms of competition.

The Chinese challenge is substantive. Moonshot's Kimi K3 release brings open weights and a new architecture to long-horizon coding, knowledge work and reasoning. The company presents frontier-level results while acknowledging that its overall evaluation still trails the strongest closed models it compares against. The release expands the set of credible alternatives for specific workloads.[2]

Alibaba's Qwen3.8 repository similarly describes the first open release of a Qwen-Max-class model and provides deployment paths for the family. DeepSeek's V4.1-Flash announcement explicitly connects architectural efficiency to lower serving costs and reduced API pricing. These are different technical and commercial approaches aimed at making capable systems more accessible.[3][4]

A buyer gains bargaining power when another system can perform the relevant work at an acceptable total cost. A developer gains options when deployment is not confined to a single provider. A model supplier faces pressure when capability that previously justified a premium becomes more widely available.

This does not require every alternative to lead every evaluation. A model can become commercially important by being sufficient for a valuable workload, easier to adapt, or economical to operate. The useful comparison includes task performance, latency, hardware requirements, licensing and the cost of the complete workflow.

Our leading expectation is a stronger division between broadly available capability and applications that earn a premium through distinctive value. For members building services, that favors portable data, reusable workflow knowledge and customer relationships over dependence on one model's temporary superiority. For buyers, it creates a reason to revisit assumptions about what must remain an expensive hosted service.

Apple is competing over where intelligence lives.

Conceptual illustration of personal computing as an alternative deployment setting for AI. — AI-generated editorial illustration. Personal computing as an alternative deployment setting for ai.

Apple's September iPhone Duo announcement is more than a new folding mechanism. It pairs the new form with A20 Pro, which Apple identifies as a 2-nanometer chip with expanded neural processing and memory bandwidth. The product connects on-device capability to an ecosystem that already controls the hardware, operating system and customer interface.[5]

The desktop side extends the argument. Apple's announced M5 Ultra Mac Studio supports configurations up to 512 GB of unified memory and 1.2 TB/s of memory bandwidth, with the 512 GB version scheduled for late October. These specifications expand the range of models and workloads that can be considered for local deployment.[6]

The strategic distinction is ownership of the execution environment. A device maker can distribute useful AI through products people already use, then connect those products to additional services when required. A cloud provider must persuade the same customer that remote execution offers enough value to justify its cost and dependence.

The contest with NVIDIA therefore reaches beyond a single chip benchmark. It involves the economics of usable memory, software support, application integration and the workloads each platform serves well. Large shared systems and local machines can have different advantages. The opportunity is to identify which work can move, and what that movement changes for the customer.

The likely near-term market is hybrid. Some tasks move onto personal or organizational hardware; others remain on hosted or clustered infrastructure. That shift can expand adoption while making a universal premium for remote inference harder to defend. Hardware and open models reinforce each other when together they make a useful local workflow practical.

Report-wide evidence · 2026-09-13

Compass Predictive Analytics

Signal gauge

100%

Independent Source Breadth

Independent Source Breadth Is 100 For The Report's Validated Factual Foundations. · Positive

tracked

Shows how many genuinely independent owners support the evidence after syndication collapse.

11 evidence references

Analytic module

module

Current Posture

Independent claim-owner cells contain 12 supporting and 0 contradicting positions.

11 evidence references

The physical supply chain can still limit the winner of a model race.

Better intelligence does not manufacture its own chips by declaration. It still needs fabrication, packaging, memory, power and deployment capacity. A lead in research becomes strategically stronger when those resources can expand alongside it.

Terafab is an explicit attempt to address that dependence. Its official presentation proposes combining logic, memory and advanced packaging for terrestrial and orbital applications. It illustrates how anticipated demand for AI and robotics can push companies toward control of manufacturing itself.[7]

The project should be followed through physical milestones: equipped facilities, qualified processes and repeat output. The same applies to other semiconductor investments. A fabrication building is one part of the system; specialized equipment, skilled personnel and workable production processes determine what it can deliver.

Orbital computing takes that chain further. Solar access is attractive, but computing hardware still needs thermal management, communications and replacement. NASA's spacecraft guidance makes heat rejection through radiators a concrete engineering requirement. Moving a facility into space changes its constraints rather than removing the need to solve them.[15]

This links back to Part II. The next advantage can belong to an organization that supplies a constraint others cannot remove quickly: useful memory capacity, reliable power, manufacturing throughput or deployment infrastructure. The model race and the infrastructure race run together, but they do not always have the same leaders. That leaves room for both powerful partnerships and unexpected challengers.

An intelligence lead can compound faster than its owner can protect it.

The strongest case for spending aggressively is the possibility of a self-reinforcing lead. Better AI could improve research and engineering, helping its owner build more capable systems and more effective industrial processes. If that cycle runs faster than competitors can respond, the gap can widen.

But protecting the knowledge behind the lead becomes part of the strategy. In January 2026, a federal jury convicted former Google engineer Linwei Ding of economic espionage and trade-secret theft involving AI infrastructure for the benefit of the PRC. The case concerned technology used to train and serve large models, placing infrastructure know-how directly inside the intelligence contest.[10]

The implication is that discovery costs and competitive benefits need not accrue to the same organization. Stolen knowledge can shorten another actor's development path. Publicly released knowledge can also diffuse lawfully, through a very different mechanism. An accurate competitive analysis must distinguish the two.

Manufacturing experience and effective deployment remain important because possession of information is not identical to the ability to execute it at scale. A country's position depends on the institutions and industries that turn knowledge into capacity, as well as the research that creates it.

Talent policy belongs in that equation. Education, research opportunity and the ability to attract and retain skilled people influence the pipeline from discovery to production. Treating legitimate international scientific mobility as espionage would damage that pipeline; ignoring documented state-backed theft would leave it exposed. The practical task is to protect valuable knowledge while keeping the research system capable of generating more of it.

Report-wide evidence · 2026-09-13

Compass Predictive Analytics

Analytic module

module

Signal Pressure Matrix

Validated independent claim-owner cells resolve to 12 support and 0 risk pressure.

11 evidence references

Analytic module

module

Evidence Density

10 source links, 11 exact spans, and 9 independent owners support report's validated factual foundations.

21 evidence references

Competition makes unilateral restraint difficult and credible coordination valuable.

Conceptual illustration of a security professional overseeing controlled access to AI systems. — AI-generated editorial illustration. A security professional overseeing controlled access to ai systems.

Amodei's proposal sets out embedded evaluators, coordination among democracies and an attempt at broader international coordination. It also explicitly recognizes that pacing must account for the competitive lead over China. That is an acknowledgment of the central strategic problem: restraint is difficult to sustain when an adversary can exploit it.[8]

Our working forecast is continued competition accompanied by stronger evaluation, access controls and efforts to coordinate restrictions. A durable unilateral pause is less consistent with the incentives than a negotiated attempt to reduce risk without surrendering advantage.

This does not require assuming that public warnings are either wholly selfless or wholly cynical. Commercial interests and genuine concerns can coexist. The revealing evidence is what a company accepts when a commitment becomes costly: independent scrutiny, unfavorable findings, meaningful restrictions and comparable obligations for the leaders themselves.

The nuclear analogy is useful as a warning about competitive incentives. States can pursue a dangerous capability because they fear the consequences of another state obtaining it first. AI adds the prospect of rapid software replication and applications across much of economic life. The comparison should illuminate that incentive, not substitute a historical image for analysis of the present technology.

Cyber defense is already part of the response. OpenAI's Codex Security applies frontier models and an agent to finding, validating and helping address software vulnerabilities. It is a concrete example of advanced AI being developed for defensive work, rather than a debate conducted entirely through hypothetical future systems.[9]

For members, the practical implication is to evaluate both capability and control when adopting powerful tools. Access to better AI can provide a real advantage. Maintaining authority over sensitive actions, data and critical systems helps keep that advantage usable.

Physical AI makes the distribution of productivity gains unavoidable.

Industrial deployment gives China a substantial starting position. The International Federation of Robotics recorded 295,000 industrial-robot installations there in 2024, representing 54 percent of the global total. That is an extensive factory footprint into which improved software and machine intelligence can be integrated.[11]

The mechanism has considerable leverage. Once a useful robotic capability is validated, its software can be distributed across compatible machines. Physical production and commissioning still take work, but each additional deployment need not repeat every step of learning from the beginning.

For a business, that can change the economics of expansion. For workers, it can change which tasks command income. The transition is likely to vary by occupation, workplace and the practicality of automating the work—not simply by whether a model can demonstrate a skill once.

Public concern is already substantial. Pew's June 2026 survey found that 71 percent of American adults expected AI to lead to fewer US jobs over the next twenty years. That measures expectations, but those expectations matter to public support for the infrastructure and policies surrounding AI.[12]

The central social question is how households gain purchasing power when more production comes from automated capital. Rising output does not settle ownership, wages or access. A society can become more productive while distributing the gains poorly.

Universal basic income belongs among the serious options for providing an economic floor. Broader ownership, shared productivity gains, accessible education and effective public services address other parts of the problem. The aim should be greater agency: people able to develop skills, care for others, create and contribute without every decision being dictated by immediate financial survival.

For members, this creates opportunities beyond replacing a worker with a machine. Integration, training, accessibility and services that help organizations and people adapt can have durable value. The strongest offering improves what people can accomplish and gives them a reason to participate in the transition.

Medical progress gives the promise of AI a human purpose.

Conceptual illustration of medical research, not evidence of a treatment outcome. — AI-generated editorial illustration. Medical research, not evidence of a treatment outcome.

Medicine supplies a concrete example. The 2012 Nobel Prize recognized the discovery that mature cells can be reprogrammed into a pluripotent state. That opened a field of research; it was not a demonstration that an adult human could simply be restored to a youthful age.[13]

The work has progressed toward clinical investigation. Life Biosciences received authorization to begin human studies of ER-100, an investigational epigenetic therapy for optic neuropathies, in 2026. The initial clinical task is to evaluate safety and tolerability in specific conditions. It is a meaningful step precisely because it asks a testable medical question.[14]

AI could contribute to discovery, design and analysis across such research. The opportunity deserves attention alongside the risks of advanced systems. A future with better treatment, greater productive capacity and wider access to useful intelligence is worth pursuing.

But prosperity has to reach people. A medical breakthrough matters through effective treatment and access. More capable machines matter through useful services, affordable goods and the distribution of the resulting gains. The success of the technology and the success of society are connected projects, not the same measurement.

Through 2028, the strongest position will combine capability, distribution and participation.

Our central outlook is that capable AI will become available through more routes while competition intensifies over the resources that make it useful at scale. Open models and efficient architectures will challenge established pricing. Integrated device ecosystems will move selected workloads closer to users. Frontier labs will continue advancing while bargaining over evaluation and restraint.

The opportunity for our members is to act where those shifts change a practical decision. Reassess expensive assumptions about model access. Preserve the ability to change providers. Build expertise and services around a real need rather than a transient model advantage. Follow the infrastructure milestones that turn a proposed capability into something customers can actually use.

The outlook would change if capable alternatives stopped improving, deployment economics prevented meaningful adoption, or enforceable international agreements substantially altered the competitive pace. These are concrete tests of the argument as the next two years unfold.

The three parts of this series lead to one conclusion. Military power needs an arsenal that can be replenished. Financial power needs commitments that create useful capacity. Intelligence leadership needs a way to turn discoveries into durable advantage—and a society able to participate in the benefits.

The goal is not merely to stay first on a leaderboard. It is to build a future in which greater capability produces greater freedom to act, for countries, institutions and the people whose lives give that progress its purpose.

Report-wide evidence · 2026-09-13

Compass Predictive Analytics

Analytic module

module

Cross Pressure

Support and risk pressure differ by 100 points.

11 evidence references

The Pivot: Compete for a future people can participate in.

Compass editorial recommendation

Compass supports continued AI development with enforceable accountability and a deliberate plan to share its gains. We should reject the choice between surrendering technological leadership and surrendering control over how that leadership is used.

Our expected course is continued competition, wider access to capable models and pressure on the resources needed to deploy them. The worse outcome is a race whose rewards concentrate while workers absorb the disruption and governments mistake unilateral restraint for defeat. That outcome can prevail without anyone choosing it as a collective goal: each institution can defend its own position while passing the consequences outward.

Start where cooperation does not require trust in an adversary's intentions. Governments should seek reciprocal incident reporting and independently tested limits on specified dangerous uses. Labs should retain responsibility for deployment decisions, support external evaluation and give customers auditable controls. Agreements should make compliance observable and violations costly. Ordinary research and useful deployment should continue; restrictions should attach to demonstrated capabilities and uses rather than serve as a blanket shield for incumbents.

The objection is serious: competitors may cheat, and verification takes time. Build defensive capacity alongside negotiation. Begin with commitments that can be checked, preserve the means to respond to violations, and widen cooperation when performance warrants it. A workable agreement must give each side a reason to comply without demanding that it abandon its future.

CEOs should make the transition concrete in the next operating plan. Choose a valuable workflow, measure the improvement, and agree with employees how the gains will support better work, paid training or compensation. Preserve the ability to change model providers. Workers should seek demonstrated competence in a real task, supported by access to tools and time to learn, rather than being told to reinvent themselves alone.

Governments should build income protection before displacement outruns the response. We favor a dependable floor alongside wider participation in productive ownership. Test income guarantees with explicit funding and purchasing-power measures; expand what works together with housing, energy and essential-service capacity. A payment cannot buy goods that the economy fails to supply.

Judge this path over the coming year by useful deployment, accountable control and gains reaching households. Rising capability with falling participation is a reason to change distribution and policy, not celebrate the leaderboard. Staying first should mean leading toward a future worth joining.

People need a way to benefit, a voice and a remedy.

At home and at work, adaptation should begin with something useful. Choose a task you understand, try an approved AI tool and check its output against what you know. Keep sensitive information out of tools that your household or employer has not agreed to use. The aim is to build judgment through experience, with enough support that learning does not become another burden carried alone.

But becoming a capable user answers only half the question. What happens when an organization uses AI to make a consequential decision about you?

When an AI system causes harm, who is responsible, and what remedy does the affected person have?

Our editorial position is that deploying an AI system should come with an identifiable responsibility to address its consequences. People should be able to learn whether AI materially influenced a decision, obtain an understandable explanation and reach someone empowered to investigate an error. A review process should explain what correction or remedy it can provide, and on what timetable. These are standards we recommend, not a claim that every institution already provides them.

For an individual facing a disputed decision, the immediate task is to establish the facts. Preserve the notice and relevant correspondence, ask the organization to identify its review process, and seek qualified help when the consequences warrant it. An automated answer that repeats the original result does not explain whether anyone examined the underlying problem.

The public discussion needs that same specificity. Compare proposals by what they require, whom they cover and how a person would use them. Consult official information about public consultations and elections; bring concrete questions to schools, employers and representatives. People can disagree about the broader direction of AI while still examining the same evidence and asking who answers when something goes wrong.

The opportunity is to make useful intelligence part of ordinary life without making ordinary people powerless before it. Progress should leave your family better able to learn, earn and make choices. That is the human test we bring to the race.

Evidence and methodology

Evidence, calculation methods and instrument scope

These instruments describe the validated factual foundations of this installment. Native Compass forecasts are unavailable for this packet; the report’s outlook and recommendations are editorial.

Evidence Reliability

95 percent Wilson lower confidence bound for the share of validated assertions carrying an exact evidence span, resolvable source URL, and validated owner-collapse identity

Evidence Freshness

mean exponential time-decay weight exp(-ln(2) * age_hours / half_life_hours) across canonical timestamped evidence at the immutable snapshot as-of

Freshness half-life: 36 hours. Timestamped sources: 7.

Independent Source Breadth

inverse Herfindahl-Hirschman concentration over validated owner-collapse assertion shares, normalized against five effective owners for the gauge value

Current Posture

validated supporting or reporting versus contradicting positions after deduplication to one cell per claim, owner-collapse identity, and pressure side

Signal Pressure Matrix

validated supporting or reporting versus contradicting positions after deduplication to one cell per claim, owner-collapse identity, and pressure side

Evidence Density

mean of bounded direct-source, exact-span, validated-assertion, and independent-owner coverage

Cross Pressure

validated supporting or reporting versus contradicting positions after deduplication to one cell per claim, owner-collapse identity, and pressure side

The Cost of Staying First | Part III: The Race Nobody Wants to Stop

Open models are changing the terms of competition.

Apple is competing over where intelligence lives.

Current Posture

The physical supply chain can still limit the winner of a model race.

An intelligence lead can compound faster than its owner can protect it.

Signal Pressure Matrix

Evidence Density

Competition makes unilateral restraint difficult and credible coordination valuable.

Physical AI makes the distribution of productivity gains unavoidable.

Medical progress gives the promise of AI a human purpose.

Through 2028, the strongest position will combine capability, distribution and participation.

Cross Pressure

The Pivot: Compete for a future people can participate in.

People need a way to benefit, a voice and a remedy.

Evidence and methodology

Evidence Reliability

Evidence Freshness

Independent Source Breadth

Current Posture

Signal Pressure Matrix

Evidence Density

Cross Pressure

Get intelligence like this delivered to your inbox

Compass Intelligence Desk

Carlos Samuels

Open models are changing the terms of competition.

Apple is competing over where intelligence lives.

Current Posture

The physical supply chain can still limit the winner of a model race.

An intelligence lead can compound faster than its owner can protect it.

Signal Pressure Matrix

Evidence Density

Competition makes unilateral restraint difficult and credible coordination valuable.

Physical AI makes the distribution of productivity gains unavoidable.

Medical progress gives the promise of AI a human purpose.

Through 2028, the strongest position will combine capability, distribution and participation.

Cross Pressure

The Pivot: Compete for a future people can participate in.

People need a way to benefit, a voice and a remedy.

Evidence and methodology

Evidence Reliability

Evidence Freshness

Independent Source Breadth

Current Posture

Signal Pressure Matrix

Evidence Density

Cross Pressure

Get intelligence like this delivered to your inbox

Compass Intelligence Desk

Carlos Samuels

The Cost of Staying First | Part II: The Balance Sheet of Power

The Cost of Staying First | Part I: The Arsenal and the Clock

Compass Weekly Nexus Pilot — September 6–12, 2026

Tech Views & AI Insights