Frontier MechanicsField report / 035
Frontier Mechanics Week 35: H20 uncertainty, Google joins MCP, Moon dust loses grip, and SSI reaches $32B
Export policy whipsawed a China-specific chip, another platform adopted an open connector, a lunar coating survived, and a pre-product lab raised at an enormous value.
Nvidia’s H20 chip appeared likely to avoid new export restrictions—for the moment. Google said it would support Anthropic’s Model Context Protocol. A dust-repellent coating from Voyager Technologies returned evidence from the lunar surface, while Safe Superintelligence reportedly raised $2 billion at a $32 billion valuation.
Each asset was valuable because of an option: continued China sales, portable tools, cleaner lunar hardware, or a future intelligence system. Options are worth more when the future is volatile. This was an exceptionally accommodating week.
Options dominated the week, but they were options of different quality. H20 could preserve access until a rule changed. MCP could preserve switching if clients honored the standard. A lunar coating could preserve mechanism life if it survived abrasion. SSI could preserve research focus by pre-funding years without product revenue. The value of an option depends on who can exercise it, what condition expires it, and whether the underlying asset improves while everyone waits.
Semiconductors
H20 lives inside the policy gap
Nvidia designed H20 as a lower-performance accelerator for China after earlier export controls blocked more capable products. Reports suggested the chip might escape additional restrictions after company lobbying, preserving a large market. Product design had become regulatory optimization: maximize useful performance while remaining below a changing threshold.
This creates short product cycles driven by Washington rather than process nodes. Chinese buyers fear future cutoff and seek domestic substitutes; Nvidia wants revenue and argues that availability keeps developers in its ecosystem. Every exemption can slow substitution and every reversal can accelerate it. Policy is now part of the chip’s specification sheet, written in erasable ink.
H20 was built after U.S. controls restricted more capable Nvidia accelerators, with reduced interconnect and performance intended to remain inside thresholds. Such products demonstrate the gap between a legal metric and a workload. A chip can fall below one peak-performance limit yet remain useful through memory capacity, software, clustering, and model optimization. Regulators respond by refining rules; vendors respond by redesigning. The feedback loop is faster than a normal silicon roadmap and slower than a policy announcement.
China represented material demand and a mature CUDA developer base. Continuing sales could generate revenue for Nvidia and keep local developers dependent on its software. A ban could deny near-term hardware while encouraging customers to port to Huawei and domestic stacks. This is the central policy tension: commercial availability can create ecosystem leverage for the exporter, and the same hardware can increase the importer's strategic capability.
Inventory makes volatility concrete. Wafers, packaging slots, boards, and customer commitments are scheduled months ahead. A policy change after production begins can create write-downs and strand supporting servers. Firms may delay shipment, accelerate orders, or design several compliance variants, all of which consume capacity. Regulatory uncertainty behaves like a tax even before a restriction takes effect.
A durable regime needs objectives that can be measured and updated on a predictable cadence. It should coordinate with allies, cover cloud access where relevant, and distinguish training, inference, and military end uses instead of relying on one marketing number. No rule will eliminate adaptation. The goal is to make strategic acquisition slower and costlier without donating the entire software market to the alternative.
H20 sits inside a policy gap created by measurable thresholds. Nvidia can reduce interconnect or arithmetic performance to remain compliant while preserving enough memory and software compatibility to retain customers. The chip's strategic value depends on how thousands combine, which workloads bottleneck on restricted features, and whether software finds substitutes. Component thresholds should be tested against complete-cluster capability.
A compliant product can still advance a domestic ecosystem through developer familiarity, CUDA dependence, and revenue. Conversely, banning it can accelerate alternative hardware and software while reducing visibility into deployed systems. Policy analysis needs a counterfactual technology path, not an assumption that denied sales equal denied compute. The objective should specify time bought and the capabilities for which that time matters.
Invoices reveal adaptation. Order size, customer type, intermediaries, and configuration can indicate whether a product serves ordinary inference or aggregation around controls. Regulators need stable guidance so distributors can investigate without freezing every sale. A constantly moving threshold produces stockpiling and redesign cycles that may enrich lawyers more reliably than they change frontier training.
AI infrastructure
Google votes for the common connector
Google joined OpenAI in announcing support for MCP, giving Anthropic’s protocol broad coverage across major model providers. Developers could build one server exposing data and tools, then connect multiple assistants. Portability weakens lock-in at the integration layer while enlarging the total tool ecosystem.
The next competition moves upward: whose client manages permissions, identity, approvals, and failures best. Standards do not end platform battles; they relocate them. USB made peripherals portable and operating systems more important. MCP may do the same for agents, assuming the security model matures before every calendar, database, and shell becomes one convenient attack surface.
Google's adoption mattered because a standard becomes real when competitors would rather interoperate than wait for their proprietary option to win. An MCP server could expose Drive files, databases, issue trackers, simulations, or internal tools to several model clients. That lets a company choose models by task and reduces duplicated connector maintenance.
Portability is incomplete unless authentication and semantics move too. One client may support eliciting user approval, streaming, or remote authorization differently from another. Tool names and schemas can be portable while policy remains vendor-specific. Enterprises need an identity layer that maps a human, agent, and server to scopes and records delegation.
A shared protocol also permits testing infrastructure. Security scanners can inspect manifests; gateways can enforce read-only modes; evaluation harnesses can replay calls across models. That ecosystem is easier to build around a stable standard than proprietary function-calling formats. The positive externality explains why even large platforms may adopt instead of own.
The standard's governance will matter as much as syntax. Version changes, extension proposals, vulnerability disclosure, and trademark control should not depend entirely on one AI vendor. A neutral foundation is not automatically necessary on day one, but credible multi-vendor stewardship prevents the common connector from becoming a strategic choke point after adoption.
A common connector matters only if implementations agree on permissions and lifecycle. Google joining MCP expands reach, but a model client still needs to discover servers, authenticate, negotiate versions, and render approval consistently. A tool that is safe in a local editor may be dangerous inside an enterprise agent with broad credentials. Conformance must include policy behavior, not merely message shape.
Vendor participation can either stabilize an open layer or turn it into the least common denominator. Extensions are inevitable for streaming, identity, rich media, and long jobs. The standard should provide capability negotiation and a portable core so proprietary features do not silently break switching. Governance needs transparent change control and test suites that no one provider can waive.
The user benefit is composability. A company can expose one inventory or design service to several model clients instead of rebuilding integrations. That leverage also centralizes risk: a vulnerable server may become reachable from every client at once. Signed packages, scoped tokens, audit logs, and revocation should arrive before the connector becomes invisible infrastructure.
Lunar materials
The Moon tries to keep the coating
Voyager Technologies said a passive electrostatic and surface-engineered coating was tested on Firefly’s Blue Ghost lander to reduce lunar-dust adhesion. Lunar regolith is sharp, abrasive, and electrically charged; it clings to seals, radiators, optics, and suits. Dust mitigation is infrastructure because nearly every surface operation touches the problem.
A coating is attractive because it adds little mass or power, but performance depends on material, exposure, and repeated abrasion. The lunar test moves the technology from simulated regolith toward real environmental evidence. Moon programs often focus on engines and landers. Long stays may be decided by whether a hinge remains willing to be a hinge.
Lunar regolith forms through micrometeorite impacts rather than water and wind, leaving sharp grains and fine particles. Solar radiation and plasma charge the surface, encouraging dust to cling electrostatically. Apollo astronauts found it abrasive, pervasive, and difficult to brush away. Future landers add engine plumes that can accelerate particles across nearby hardware.
A passive coating changes surface energy or conductivity so particles adhere less strongly. It consumes no operational power and can cover complex shapes. Performance is surface-specific: optical windows, radiators, seals, fabrics, and solar cells have different requirements. A coating that sheds dust but changes emissivity or degrades under ultraviolet radiation can move the problem.
The Blue Ghost exposure supplied actual regolith, radiation, vacuum, thermal cycles, and local charging conditions that simulants only approximate. Interpreting the result requires a control surface, before-and-after imagery or measurements, exposure duration, and knowledge of how dust arrived. One clean coupon is a qualification step, not proof of years of operation.
Dust mitigation will likely be layered. Landing pads and plume models reduce deposition; passive coatings reduce adhesion; electrodynamic screens move charged particles; seals and covers protect critical interfaces; cleaning tools address accumulation. The coating is valuable precisely because it can reduce load on every active layer. Small materials science can create system margin across a base.
Lunar dust attacks coatings through abrasion, adhesion, charging, and optical contamination. Sharp regolith grains can scratch surfaces; electrostatic forces make particles cling in vacuum; thermal cycling changes bonds; and ultraviolet exposure alters materials. A coating that repels a clean simulant on Earth may fail after radiation and repeated mechanical contact. Testing must reproduce the coupled surface environment.
The metric depends on the surface. A camera needs optical transmission and stray-light control; a radiator needs emissivity; a seal needs particle exclusion; a solar array needs power retention. One universal contact angle or dust-removal percentage is not enough. Engineers should measure performance after representative deposition, cycling, cleaning, and wear, then compare the coating with operational alternatives such as covers or electrodynamic screens.
Maintenance is part of the habitat architecture. Dust carried through an airlock can affect lungs, mechanisms, connectors, and filters long after an excursion. Passive coatings reduce burden but cannot own the whole hazard. Suitports, cleaning zones, material selection, and monitoring form a layered control. The Moon keeps the coating honest because failure accumulates across missions rather than disappearing after a laboratory photograph.
Venture capital
SSI sells the undistracted laboratory
Safe Superintelligence, founded by Ilya Sutskever and colleagues, reportedly raised $2 billion at a $32 billion valuation despite having no public product. The pitch is organizational focus: build a safe superintelligent system without near-term product pressure. Investors receive a claim on an unusually scarce team and an enormous outcome.
Avoiding products can protect research from customer deadlines and removes revenue that would otherwise fund compute. The company therefore needs capital and trust in advance. The valuation is less a discounted cash-flow model than a market price for talent concentration and strategic optionality. It may be rational. It is also what a very expensive promise looks like before evidence interrupts.
SSI was designed around one goal—safe superintelligence—and explicitly avoided product distraction. The structure can protect researchers from quarterly feature work and revenue incentives that reward deployment before understanding. It also removes the feedback and discipline that customers provide. A lab without products must define intermediate evidence rigorous enough to prevent mission language from becoming insulation against evaluation.
A $32 billion post-money value does not imply $32 billion of current assets or conventional forecast cash flow. It reflects investor belief in founder talent, hiring concentration, compute access, and a winner-take-most future. The payoff distribution is extreme: a successful system could be worth vastly more, while a safety-first research approach with no product may return little. Venture pricing can rationally tolerate that variance for a small portfolio allocation.
Capital buys time and compute but also creates expectations. Even patient investors have fund lives, ownership goals, and competitive pressure. Raising a large round early can lock the lab into expensive scale and make a smaller but scientifically sound result feel inadequate. Independence from product deadlines is not independence from capital markets.
The safety claim should become observable through research practices: threat models, evaluations, access controls, governance, staged deployment decisions, and willingness to publish negative results. Secrecy may protect capabilities and makes external trust harder. SSI's organizational experiment is whether concentrated private capital can support caution without public accountability or customer feedback. The answer will be institutional before it is technical.
SSI's undistracted-laboratory thesis treats organizational attention as a scarce input. A lab insulated from product launches can pursue long research arcs and accept negative results. The benefit is real when the objective requires coherent architecture rather than quarterly feature pressure. The risk is that absence of customers removes reality checks and lets internally elegant metrics drift away from useful or safe capability.
Compute and talent still impose milestones even if revenue does not. Investors should ask what empirical uncertainty each tranche resolves: scaling behavior, data efficiency, controllability, or an architectural prototype. Secrecy can protect focus but also prevents external falsification. A credible research company can publish methods and failure lessons without revealing the asset it believes confers advantage.
Safety as the sole product goal needs operational content. It could mean alignment under capability growth, secure training, interpretability, bounded autonomy, or governance. Each implies different experiments and threat models. SSI will deserve patience if it converts the mandate into tests that could prove its approach wrong. Undistracted work is an input; disciplined exposure to disconfirming evidence is the output.
H20, MCP, lunar coatings, and SSI all derive value from preserving future action. Sell into China, switch model clients, keep a mechanism moving, pursue research without revenue. Optionality is useful and expensive because it asks somebody else to bear uncertainty now. This week, markets were unusually happy to volunteer.
An option is valuable because it avoids commitment while preserving upside. H20 let Nvidia retain a market without shipping its best part. MCP let developers retain model choice. A coating let lunar systems avoid heavier active cleaning. SSI let researchers postpone product obligations. Each option also left someone else carrying risk.
Chinese customers faced cutoff, tool users faced a broader security surface, lunar programs faced incomplete qualification, and SSI's investors and future public faced limited evidence. Options are not free choices suspended in air. They are contracts over uncertainty, with exercise rights distributed unevenly.
The best way to value them is with gates. Which rule cancels sales? Which protocol behavior proves portability? Which abrasion exposure ends coating qualification? Which research evidence justifies another compute tranche? Explicit gates keep optionality from becoming an excuse to defer judgment indefinitely.
Week 35's valuations were generous because futures were wide. Technical progress will narrow them. That is the point of the experiments—and the moment when narrative option value must surrender to measured performance.