Frontier MechanicsField report / 034
Frontier Mechanics Week 34: Llama 4’s mixture, Gemini’s price, Isar’s crash, and Aetherflux’s round
Meta routed tokens through experts, Google priced a premium model, Europe’s new orbital rocket flew briefly, and orbital solar collected $50 million.
Meta released Llama 4 with mixture-of-experts architectures and controversial benchmark presentation. Google priced Gemini 2.5 Pro at the top of its model range. Isar Aerospace’s Spectrum rocket flew for roughly half a minute before crashing, and Aetherflux raised $50 million for a space-solar demonstration.
The week was about expensive first versions: sparse models that still require trust, premium inference that still requires routing, a brief flight that still produced data, and power beaming that still requires an orbital proof.
Sparse computation, premium inference, a thirty-second flight, and a future power beam all asked readers to distinguish a partial system from a completed one. Llama 4 activated only part of its weights. Gemini's token price omitted application retries and routing. Spectrum validated several phases before failing the mission. Aetherflux proposed to measure a chain whose individual efficiencies multiply. The denominator was the story each time.
Open models
Llama routes tokens—and expectations
Llama 4 Scout and Maverick used mixture-of-experts designs, activating a subset of parameters for each token. This expands total model capacity without paying the full compute cost on every step. Meta promoted strong benchmark results, but criticism followed because a tuned experimental version appeared on a popular leaderboard rather than the plain released model.
Mixture routing is technically efficient; benchmark routing is less charming. Open models create trust through inspectability, so presentation matters as much as weights. If a score depends on a special chat-tuned variant, users need to know. The broader lesson is that evaluation has become part of model governance, not a detachable marketing chart.
A mixture-of-experts transformer replaces some dense feed-forward blocks with multiple expert networks and a router. For each token, the router selects a small subset, so total parameter capacity can rise without activating the entire model. The design can learn specialization and improve compute efficiency. It also creates load-balancing problems: popular experts become bottlenecks, rarely chosen experts undertrain, and routing traffic stresses accelerators and networks.
The number printed in a model name therefore needs qualification. Total parameters describe stored capacity; active parameters approximate per-token arithmetic; memory and communication depend on quantization, batching, expert placement, and cache. A sparse model with a huge total may serve cheaply only when infrastructure keeps experts local and batches routes efficiently. Consumer hardware often cannot hold all experts even if each token uses few.
Meta's benchmark dispute mattered because open releases depend on reproducibility as a substitute for vendor trust. If a leaderboard submission uses a separately tuned system prompt or unreleased variant, the public weights do not inherit the score. Small post-training differences can move chat preference dramatically. A model card should tie each result to an exact checkpoint, prompt template, tool setting, and evaluator.
Open weights still created real value: labs could inspect routing, fine-tune domains, run models privately, and compare claims. The episode did not negate the architecture. It illustrated that openness is a chain, not a binary label. Weights without training data, evaluation code, or deployment assumptions expose some layers and leave others as marketing.
Mixture-of-experts systems route each token through selected parameter blocks, raising total capacity without activating the whole model. The router must balance specialization and load. If a few experts receive most tokens, they bottleneck devices while others sit idle; auxiliary losses can encourage balance but may reduce semantic specialization. Serving performance depends on communication across accelerators as much as nominal FLOPs.
Expectation routing is the product analogue. Meta can market a large parameter count while users experience active parameters, context limits, quantization, and safety layers. Comparisons should state activated compute and memory traffic per token alongside total weights. Otherwise a sparse model can appear simultaneously enormous for capability and small for cost, choosing whichever denominator flatters the claim.
Open access improves scrutiny if weights and enough implementation detail are released. Researchers can test routing collapse, multilingual expert allocation, fine-tuning stability, and vulnerability to inputs that overload specific experts. The architecture's merit is efficient specialization under real traffic. A sparse checkpoint that requires heroic networking or loses quality after quantization may be less economical than its active-parameter count suggests.
Model economics
Gemini 2.5 prices the premium thought
Google announced paid API pricing for Gemini 2.5 Pro, making it the company’s most expensive model. Long context and reasoning consume memory and computation in ways a simple per-token number can obscure. Developers pay for input, output, cache behavior, and the possibility that the model thinks longer.
Premium pricing is defensible where the model replaces valuable work, but application margins depend on controlling surprise. Teams will cache context, route easy tasks downward, and reserve 2.5 Pro for failures or high-value steps. The model market is beginning to resemble cloud storage tiers: intelligence differentiated by latency, depth, and invoice anxiety.
API prices list input and output tokens, but reasoning models may consume hidden internal tokens and long-context requests occupy memory while they run. An application also pays for retrieval, vector storage, tool calls, guardrails, logging, and failed generations. The marginal cost of a successful customer outcome can be several times the model line item.
Prompt caching changes the curve for long repeated context. A legal corpus, codebase prefix, or product manual can be processed once and reused while cache remains valid. That rewards stable prefixes and careful context construction. Teams that paste the same megabyte into every request without cache discipline can erase the value of a more accurate model through memory cost.
Routing is the economic control plane. A small model can classify intent, answer simple work, or construct a first draft; Gemini 2.5 Pro can handle escalations and high-value steps. Verification may use deterministic tests or a second model. The router should optimize expected total cost of correctness, including human review and customer harm, rather than chase the lowest token price.
Premium pricing also reveals willingness to pay for latency and integration. A model that fixes a production bug or completes a complex analysis can be cheap at dollars per request. The same model drafting thousands of low-value support replies can be ruinous. 'Too expensive' is not a model property. It is a mismatch between allocation and the economic value of the step.
Premium thought is a scheduling promise. Longer reasoning can occupy an accelerator sequentially, delaying other requests even when average throughput looks acceptable. The service needs admission control, priority, and early stopping so a few pathological prompts do not dominate tail latency. Pricing should reflect reserved capacity or actual computation with a ceiling users can set.
More tokens are not necessarily more search. A model can restate assumptions and reinforce an early error. Useful effort changes representation, checks with a tool, explores a contrary hypothesis, or runs a verifier. Evaluation logs should classify what the additional budget did and measure marginal accuracy at each step. That reveals the point at which thought becomes ceremony.
For enterprise workflows, reproducibility matters more than a charismatic maximum. A versioned effort policy and recorded tool trace let teams validate outputs against a cost envelope. If Google changes routing to manage fleet load, customers need to know. Premium reasoning becomes infrastructure when quality and latency distributions can be contracted, not when an interface displays a thinking animation.
European launch
Spectrum earns thirty seconds
Isar Aerospace launched Spectrum from Norway on Europe’s first orbital attempt by a commercial rocket from continental soil. The vehicle lifted off, lost control, and fell into the sea after about thirty seconds. It did not reach the intended trajectory; it did validate pad operations, ignition, and a real flight environment.
Calling every failed maiden flight a success is public-relations anesthesia. Calling it worthless ignores how rockets are developed. The useful question is whether telemetry identifies correctable causes and the company can finance another attempt. Europe wants independent commercial launch capacity. Independence, inconveniently, requires paying for the flights that prove it does not yet have it.
Spectrum was a two-stage liquid rocket designed for small and medium payloads, launching from Andøya in Norway. Its brief maiden flight verified countdown, fueling, engine ignition, hold-down release, pad interfaces, telemetry, and the initial aerodynamic regime. Those are real milestones. The mission objective was orbit, and the vehicle did not approach it. Both sentences belong in the same report.
Loss of control shortly after liftoff focuses investigation on propulsion symmetry, thrust-vector control, inertial measurement, guidance commands, structural response, and software. Video can suggest pitch or yaw, but telemetry determines whether the vehicle commanded the motion, could not execute the command, or received a bad state estimate. Recovery depends on preserving data and reproducing the fault on test hardware.
A maiden flight is expensive integration testing because rockets contain interactions that stands cannot fully duplicate. Cryogenic propellant sloshes, flexible structures bend, winds change, and engines vibrate together. Still, 'we learned' should be followed by a closed corrective-action list and evidence on the next flight. Learning that does not alter a controlled configuration is public relations.
Europe's strategic objective gives Isar access to patient customers and scrutiny. Sovereign launch matters only if vehicles fly often enough to serve missions. Public support can bridge early failures; it should be milestone-based and preserve competition among launchers and spaceports. Independence purchased through one protected provider can reproduce the dependency it was meant to solve.
Thirty seconds of spectrum can be more scarce than a satellite because the transmission must coexist with incumbents, national allocations, and interference geometry. A demonstration window tests coordination and waveform compatibility. Continuous service requires rights across territories and enough link margin to operate without monopolizing the band. Regulatory timing is part of network capacity.
Spectrum efficiency depends on modulation, coding, antenna gain, beam reuse, power, and terminal capability. Peak throughput in a clean test does not describe edge-of-beam users or rain fade. Operators should report delivered bits per hertz per beam under concurrent load and the interference assumptions used. Those numbers determine whether the spectrum asset supports a business rather than a press release.
Secondary markets and sharing can improve utilization if enforcement is fast. Dynamic coordination databases, geofencing, and telemetry can allocate access, but failures may disrupt safety-critical incumbents. A company earns trust by making emissions observable and shutdown deterministic. The scarce resource is not an abstract frequency; it is predictable permission to radiate at a place and time.
Energy
Aetherflux funds the orbital receipt
Aetherflux raised $50 million to launch a demonstration that would collect solar power and transmit it by infrared laser. The first target is small relative to grid generation. Its purpose is to close the technical loop: pointing, conversion efficiency, atmospheric propagation, receiver performance, and operational safety.
A demonstration also changes fundraising. Investors can debate measured losses instead of a conceptual architecture. Defense customers may tolerate high costs for power delivered to remote areas; that beachhead could finance larger systems. The mission is designed to produce a number—how many useful watts arrive—which is exactly the kind of thing a grand idea eventually needs.
A space-solar power chain begins with incident sunlight, converts it to electricity, runs a laser, propagates through pointing error and atmosphere, and reconverts received light into useful electricity. If each stage is 70 percent efficient, four stages deliver about 24 percent before operational losses. The system can still be useful if orbital illumination, remote access, or avoided fuel has high value. It cannot be evaluated by panel output alone.
Infrared laser beaming trades receiver size and directionality against clouds, absorption, eye safety, aircraft coordination, and pointing. A narrow beam reduces geometric loss but requires precise acquisition and tracking between a fast satellite and ground target. Atmospheric turbulence can spread and wander the spot. A pilot demonstration should measure delivered power across weather and off-nominal pointing, not only a clear-sky peak.
Low Earth orbit limits continuous dwell over one site. A small spacecraft passes quickly, so useful service needs storage, multiple satellites, or customers who value intermittent bursts. Higher orbits improve access and worsen distance loss and launch cost. Defense customers might value power to sensors or forward positions where fuel logistics dominate, giving the system a premium niche before grid-scale ambition.
Safety and licensing are system components. The beam needs interlocks, no-fire zones, aircraft detection, predictable shutdown, cyber protection, and public reporting of wavelength and intensity. Aetherflux's first mission could make the debate quantitative by publishing end-to-end efficiency, spot control, uptime, and fault response. A civilization-scale pitch becomes credible one measured watt at a time.
Orbital solar power begins with an unfavorable chain of efficiencies: sunlight collection, electrical conversion, radio-frequency or optical conversion, transmission, atmospheric loss, receiver conversion, and grid conditioning. Multiplying optimistic values can still yield substantial loss, while launch mass and pointing hardware add capital cost. Aetherflux's small demonstration should measure each conversion separately so a weak link cannot hide behind end-to-end ambiguity.
Microwave transmission requires aperture and safety tradeoffs. Diffraction spreads a beam according to wavelength and transmitter size; a tighter spot needs a larger orbital antenna or shorter wavelength. Ground receivers occupy land and must coexist with aircraft, communications, and people. Beam steering and immediate shutdown on pointing error are central safety functions, not secondary software features.
The strategic niche may be remote or contested users rather than a terrestrial grid. A customer with expensive diesel logistics may value modest continuous power at a price no utility would accept. The receipt should therefore include watts at the load, duty cycle, receiver area, weather, and total launched mass. Beamed power becomes credible one audited energy ledger at a time.
Llama 4, Gemini pricing, Spectrum, and Aetherflux all ask to be judged as systems rather than isolated numbers. Activated parameters are not total parameters; token prices are not application costs; flight duration is not learning; collected sunlight is not delivered power. Frontier marketing selects the flattering denominator. Frontier analysis puts the rest back.
Partial success is not a euphemism when the parts are named. Sparse activation can lower compute while routing remains hard. A premium model can improve results while workload cost remains uncertain. A failed rocket can validate ground and flight phases while orbit remains unproven. A solar demonstration can close pointing without approaching grid economics.
The discipline is to maintain a requirements tree. Each claim maps to an observable test, each test maps to evidence, and unmet objectives remain visible. Frontier programs get into trouble when one successful branch is narrated as the root requirement.
Investors can fund intermediate evidence without pretending it is revenue. Engineers can celebrate a subsystem without calling the mission complete. Users can value a model without accepting every leaderboard chart. This is not pessimism. It is how staged risk becomes financeable.
Week 34 rewarded systems that made the next unknown smaller. The ones that survive will preserve that accounting even when a more flattering denominator is available.