Valley of Death by Sharon Weinberger

BOOK REVIEW

Silicon Valley Goes to War, and Brings a Pitch Deck

Valley of Death  by Sharon Weinberger

Little, Brown and Company  ·  September 2026

In 2013, about ten scientists and strategists gathered in a Virginia home to play a game whose premise sounds like the pitch for an airport thriller: could a rich man build and sell nuclear weapons? The answer, reached with nothing more than open-source research and shopping on Alibaba, was yes. It would take roughly a billion dollars, about five years and four nominally unrelated companies spread from Zurich to Kazakhstan to Dubai, and it would be hard for anyone to stop. The Pentagon had commissioned the exercise, and by one participant's account the people involved were told to scrub their computers afterward, lest the study double as a how-to manual. Sharon Weinberger opens Valley of Death with this war game, and it is a shrewd choice: a lurid, memorable image that carries her argument before she has made it.

The argument is that the government's old monopoly on the machinery of war is passing to a small band of technologists and venture capitalists who have their own politics, their own balance sheets and a deep suspicion of the institutions they sell to. She calls the new arrangement the Pentagon-Silicon Complex, a deliberate echo of the military-industrial complex that Eisenhower warned about in 1961. The title borrows an industry cliché, the “valley of death” between a promising prototype and a paying military customer. The book's real subject is what happens now that the valley is being crossed. The numbers she offers are startling: by her account, venture capital poured nearly $50 billion into defense startups in 2025, more than double the previous year.

Weinberger, a veteran national security reporter who spent years at the Wall Street Journal and wrote an earlier book on DARPA, is at her best when she goes backward in time. The first chapter is a small marvel. It begins with a teenage mathematician recruited by the CIA in 1969 and arrives, improbably, at Steve Jobs, whose struggling NeXT computer company was kept afloat by spy-agency orders. The machines were expensive, elegant and almost unsellable to ordinary customers; the intelligence community bought them by the thousands, and a billionaire investor leaned on Jobs when he hesitated. When Apple bought NeXT in 1996, the operating system that came with it quietly shaped the company's future. The anecdote punctures one of Silicon Valley's favorite myths, that it is a civilian garden of garage tinkerers and counterculture dreamers, and it establishes the book's governing irony: the Valley was grown on Pentagon money long before it started pretending otherwise.

From there the book becomes a group portrait, and the cast is formidable. Elon Musk, who wanted to go to Mars and found that the Pentagon would help pay for the rockets. Peter Thiel, who told a Stanford class that competition is for losers and meant it as a business plan, not a provocation. Alex Karp, Thiel's partner in Palantir, a company named after the seeing stones of Tolkien and born from the failures of intelligence sharing that preceded September 11. Palmer Luckey, the virtual-reality prodigy turned weapons maker. Weinberger sketches them with a reporter's economy, usually through a telling scene instead of a character study, and she is careful to let their ambitions speak for themselves. Thiel's monopoly gospel, in particular, reads very differently once you notice how closely SpaceX's grip on military launches, and Palantir's own ambitions, resemble it.

The book is sharpest when it deflates its subjects. Take Project Maven, the Pentagon's early attempt to use artificial intelligence on drone footage, remembered mostly for the employee revolt that drove Google away from it. Weinberger shows that the rebellion mattered less than anyone thought. The system's first field test went badly enough that, by a Palantir executive's own telling, it struggled to tell people from sheep. Its architects concede that the Pentagon lacked the basic digital plumbing, data pipelines and cloud infrastructure that any serious AI program needs. What Maven really did, she argues, was open the floodgates, because a wave of companies saw that working with the military was now respectable and lucrative, particularly as the Chinese market closed to them. It is a nicely counterintuitive account, and it is typical of her method: take the official story, find the people who lived it, and let the gap between claim and reality do the work.

Politics enters in force in the middle chapters, and Weinberger handles it with unusual restraint. Thiel's endorsement of Trump at the 2016 Republican convention was a shock in a Valley that largely recoiled from the candidate. The fight over JEDI, a $10 billion Pentagon cloud contract, turned into a proxy for Trump's feud with Jeff Bezos, and the president needed a moment just to identify the contract when a reporter asked him about it. Palantir went public in 2020 and was worth more than $20 billion by the end of its first day of trading. Then Biden's election, perversely, gave the industry something Trump could not: legitimacy among the left-leaning engineers whose labor it needed. The book is good on the way a movement that began as an embarrassment became, within a few years, a career path with a fan base.

The war chapters are where the title's anxiety turns concrete. Russia's invasion of Ukraine made weapons-building fashionable, and Starlink terminals, commercial satellite imagery and Amazon's cloud became instruments of a real war. There is a wonderful, absurd image of Palantir's chief executive doing tai chi near the Polish border with Ukraine, drawing suspicious looks from guards. Weinberger turns to a harder question in her conversation with Luckey about Musk's control over Starlink in Ukraine: if the company's argument is that elected governments, not executives, should set foreign policy, what happens when elections put those executives in power? Luckey never quite answers, and her refusal to let him off is the best moment of interviewing in the book. The same instinct drives her account of Gaza, where Palantir announced that it stood with Israel, an unusual political stance for a defense contractor, and she asks plainly why a company would choose to.

Trump's second victory is where the book arrives at its most chilling formulation. Karp appears at the Reagan National Defense Forum in December 2024, practically giddy, attacking Berkeley, the United Nations and Democrats, and promising that the company's customers will win everything. Musk answered the speech online with a single word, “Based.” Weinberger lets the scene sit, observing that no one in the room, including Democratic lawmakers, contradicted him, and she ends the chapter with a sentence that earns its bluntness: the war bros were no longer just lobbying the government, they were in it. The next chapter opens on the tech leaders' dinner with Trump in September 2025, a ceremony of flattery that she rescues from caricature by quoting a venture capitalist's dry remark that corporate power and state power have simply replaced the royal court.

For all its virtues, Valley of Death has real weaknesses, and they are mostly structural. The book is episodic by design, a chain of vivid scenes held together by theme more than by narrative momentum, and a reader without a specialist's familiarity with defense contracting may drown in the names of companies, programs, offices and investors. Palantir, Karp and the Thiel network occupy so much of the foreground that other actors, including the traditional primes the new firms claim to displace, recede into the background. And the book's central metaphor is slipperier than it first appears. Weinberger concedes that these founders probably won't build nuclear weapons; the real worry is autonomous drones, AI-assisted targeting and software that quietly shapes decisions. The billionaire's bomb is an arresting frame, but it promises a danger the book then relocates, and the shift is not always clean.

That said, the book's most admirable quality is its refusal to be a polemic. Weinberger is wary of the concentration of power in a few hands, but she is equally wary of the industry's critics who assume these companies are competent. Her deepest skepticism is reserved for the proposition that the new weapons work. A venture investor tells her, in effect, that it does not matter to him whether what he funds works. A drone company's chief executive explains, over wine at a Northern California vineyard, that he wants to own the first twelve choices on Amazon's list so that customers feel they are choosing freely. The firm, valued at more than $3 billion against modest revenue, appeared in a federal database to hold one government contract worth $167,000. A General Atomics spokesperson went so far as to call Anduril the Theranos of defense, a comparison Weinberger notes is unfair, since no one has alleged fraud, but she lingers on it because the rhetoric, hype and valuation logic are familiar.

The epilogue turns to the scenario that justifies the whole enterprise: a Chinese invasion of Taiwan. Weinberger describes tabletop war games in which the United States and its allies usually win but at a terrible price, with two aircraft carriers and thousands of sailors lost. She sets that grim arithmetic beside a Palantir commercial that depicts a nearly bloodless war fought by drones and a commander conducting them like an orchestra, a vision of warfare that, she points out, does not yet exist and that Palantir is not building. Investors, in her phrase, were betting on vibes. Michael Burry's decision to short the stock and the company's decline of about a quarter from its late-2025 high, as of March 2026, serve as a reminder that the market's enthusiasm may be running ahead of the evidence. The risk, she suggests, cuts both ways: a bust could drag down a whole class of startups just as the Pentagon has learned to depend on them.

She closes with a conversation with Will Roper, the official behind Maven, who urges her to end on the hundreds of companies caravanning across the valley of death and the possibility that the experiment fails if no one welcomes them on the other side. Weinberger chooses a slightly different ending. The danger, she says, is not that Silicon Valley will fail to deliver, but that a few firms will succeed so thoroughly that they displace the old primes and carry their founders' convictions into the heart of national security. She returns to Eisenhower, who understood the need for the industrial development he was warning about and urged the country not to fail to grasp its grave implications. It is a restrained and fitting conclusion, and it leaves the reader with a question rather than a verdict.

Valley of Death is not a perfect book. It is crowded, occasionally repetitive and more convincing as reportage than as prophecy. But it asks the right question at the right moment, and it asks it with a rare combination of access, skepticism and fairness. The pleasures here are the pleasures of good journalism: the unexpected anecdote, the telling detail, the interview in which the subject says slightly more than he meant to. For anyone trying to understand who will build, sell and ultimately control the next generation of weapons, and what that might do to a democracy that has long assumed it holds the reins, this is essential reading. Read it for the scenes. Stay for the unease.

Review of Shoe Dog: A Memoir by the Creator of Nike

1. Introduction

Shoе Dog: A Memoir by the Creator of Nike is an autobiographical account written by Phil Knight, the founder of Nike. In this book, Knight describes his journey from having a small idea involving athletic shoes to building one of the most recognized sports brands in the world. Rather than presenting success as a simple or predictable process, Knight focuses on the difficulties, risks, mistakes, and personal experiences that shaped his entrepreneurial journey. The memoir provides a human perspective on entrepreneurship and demonstrates that building a successful company requires persistence, adaptation, and the ability to continue despite uncertainty.

2. Context and Objective

The story takes place primarily during the 1960s and 1970s in the United States, a period in which the athletic footwear industry was developing significantly. Knight studied at the University of Oregon and later attended Stanford Business School. After completing his studies and serving in the military, he traveled and developed the idea of importing Japanese athletic shoes into the United States.

His first company, Blue Ribbon Sports, was created with the support of his former coach, Bill Bowerman. The company initially focused on selling athletic shoes to runners. However, the business faced numerous financial and operational difficulties, including problems with financing, suppliers, banks, and competitors. Eventually, these challenges contributed to the transformation of Blue Ribbon Sports into Nike.

The objective of the book is not to provide a conventional business manual. Instead, Knight offers a personal account of the experiences, decisions, relationships, and difficulties involved in building his company.

3. Summary of the Work

The memoir begins with Knight reflecting on his uncertainty about his future and the experiences that influenced his entrepreneurial ambitions. His travels and education helped him develop the idea of importing Japanese running shoes, which eventually became the foundation of Blue Ribbon Sports.

The company began on a relatively small scale, with Knight and his partners selling shoes directly to runners. One of the most important figures in the story is Bill Bowerman, Knight's former track coach, who became both a business partner and an important contributor to the development of athletic footwear.

As the company expanded, it encountered significant financial and organizational challenges. Maintaining sufficient cash flow, obtaining financing, dealing with suppliers, and competing in a growing market became constant concerns. Knight also describes the importance of the people who worked with him and the relationships that helped the company survive difficult periods.

Eventually, Blue Ribbon Sports developed into Nike. The creation of the Nike name and the famous Swoosh symbol represented an important stage in the company's transformation. As the company grew, however, new challenges emerged, demonstrating that growth did not eliminate uncertainty but instead created different and more complex problems.

4. Critical Analysis

One of the main strengths of Shoe Dog is its human approach to entrepreneurship. Knight does not portray himself simply as a successful businessman. Instead, he describes his doubts, mistakes, financial concerns, conflicts, and moments of uncertainty. This makes the story more realistic and allows readers to understand that entrepreneurial success is often the result of a long and uncertain process.

Another important strength is the attention given to the people behind the company. Figures such as Bill Bowerman and the members of Knight's early team played an essential role in the development of the business. The memoir therefore presents Nike not only as the product of one individual's vision but also as the result of collaboration, relationships, and shared effort.

The book also effectively connects sports with business. Knight's personal connection to running influenced his understanding of athletes and athletic footwear. This connection contributed to the development of a company whose identity became strongly associated with sport and performance.

However, the book also has a limitation inherent in autobiographical works: the story is presented primarily from Knight's personal perspective. His memories and interpretations shape the way events and other people are represented, meaning that other participants might interpret some situations differently. In addition, some sections may feel longer than necessary for readers interested exclusively in business strategy. Nevertheless, the personal details contribute to a better understanding of Knight's motivations and the culture that developed within the company.

5. Conclusion

Shoe Dog goes beyond being a traditional corporate history. It presents the development of Nike as a long and uncertain process characterized by financial difficulties, mistakes, conflicts, personal relationships, and important decisions. The book demonstrates that successful companies are not necessarily created through a guaranteed plan, but can emerge through persistence, adaptation, teamwork, and continuous problem-solving.

The memoir is particularly relevant for students and professionals interested in management, entrepreneurship, marketing, business, and leadership. It can also appeal to general readers because of its personal and narrative style.

Ultimately, Knight's story shows that behind a globally recognized company was once a small and uncertain business project. Its development depended not only on an idea, but also on the people who supported it and the willingness to continue working through years of challenges.

Artificial Intelligence Computing in Low Earth Orbit

 Artificial Intelligence Computing in Low Earth Orbit: Silicon Survival in the Space Environment and Its Consequences for AI Deployment

 Abstract

In 2026, at least eight companies — including Google, Nvidia, SpaceX, Starcloud, Amazon and Blue Origin — are competing to place artificial intelligence compute capacity in low Earth orbit (LEO), driven by electricity shortages and growing regulatory resistance to terrestrial data centers. Starcloud has already run Google's Gemma language model on an Nvidia H100 GPU in orbit, and Google placed four tensor processing units (TPUs) into orbit aboard an experimental satellite on October 1, 2026, as part of Project Suncatcher. Yet silicon designed for climate-controlled rooms on Earth's surface must now operate in a radically different environment: ionizing radiation without atmospheric or sufficient magnetic shielding, extreme thermal cycling, a vacuum that eliminates convective cooling, and constant exposure to micrometeoroid and orbital debris impacts. This paper reviews the technical survival requirements — radiation hardening, architectural redundancy, radiative thermal management and software-level fault tolerance — needed to prevent malfunction of AI accelerators in orbit, and examines how these constraints shape which applications are actually viable: low-latency geospatial inference and constellation autonomy, rather than large-scale foundation-model training.

1. Introduction

Demand for compute to train and run artificial intelligence models has outpaced the ability of terrestrial power grids to supply it. In Virginia, Ireland and other traditional data-center hubs, operators face moratoria, multi-year grid interconnection queues, and mounting community opposition over water and electricity consumption. Facing this bottleneck, a growing number of companies have turned their attention to low Earth orbit, where a satellite in a sun-synchronous orbit can receive near-continuous sunlight and, according to Google's estimates, generate up to eight times more solar power per panel than an equivalent array on the ground.

The idea itself is not new — communications and Earth-observation satellites have operated electronics in orbit for decades — but the scale and type of silicon now being proposed are. Where traditional satellites use radiation-hardened ("rad-hard") components built on older, extensively validated fabrication processes, the new wave of orbital AI projects aims to fly high-performance accelerators — GPUs and TPUs originally designed for terrestrial data centers — with minimal modification. Starcloud placed an unshielded Nvidia H100 GPU into orbit in November 2025 and, a month later, successfully queried Google's Gemma model running on it. Nvidia, for its part, unveiled a line of computing platforms purpose-built for orbital data centers at its 2026 GTC conference, along with a hardened module based on its Vera Rubin architecture slated for later in the decade. Google, SpaceX and Amazon have announced parallel efforts built around proprietary TPUs, a Starmind architecture, and AWS Outposts hardware, respectively.

This paper examines, from a technical and forward-looking perspective, two related questions: what must be guaranteed for these chips to survive and operate reliably in an environment far more hostile than Earth's surface, and what consequences these survival requirements carry for the kinds of AI applications that can realistically be deployed in orbit over the coming years.

2. The Space Environment as an Engineering Problem

Four environmental factors set low Earth orbit apart from any data center built on the ground.

Ionizing radiation. Without the atmosphere or most of the magnetic shielding that protect surface electronics, orbiting circuits are exposed to high-energy charged particles from the solar wind, the Van Allen belts and galactic cosmic rays. These particles cause two kinds of damage. The first, called a single-event effect, occurs when an individual particle flips the state of a memory bit or triggers a transient short circuit in a transistor; it is a probabilistic phenomenon that can silently corrupt an inference calculation or, in the worst case, cause a destructive latch-up that disables the chip. The second, total ionizing dose damage, progressively degrades the electrical properties of transistors over the course of a mission until the device no longer meets specification.

Thermal cycling and vacuum. In the absence of air, an orbital data center cannot dissipate heat through convection, the mechanism nearly all terrestrial cooling systems rely on. The only available mechanism is thermal radiation into deep space, far less efficient per unit of surface area. Compounding this, a satellite in low orbit passes through Earth's shadow several times a day, subjecting electronics to repeated temperature swings of tens of degrees within minutes — a thermal fatigue regime very different from that of a temperature-controlled server room.

Vacuum and outgassing. The vacuum of space causes certain materials common in terrestrial electronics — adhesives, polymer insulators, some coatings — to release trapped gases (outgassing), which can deposit residue on sensitive optical or electronic surfaces and alter the dielectric properties of components.

Micrometeoroids and orbital debris. Low Earth orbit is increasingly crowded with satellite fragments, spent rocket stages and natural particles traveling at several kilometers per second. An impact, even from a millimeter-scale fragment, can puncture solar panels, radiators or the housing of a compute module.

None of these four factors has a direct equivalent in terrestrial data-center design, and each demands a distinct engineering response.

3. Survival Requirements to Prevent Malfunction

Radiation hardening and design-level mitigation. Two complementary strategies exist. The first is to fabricate the chip itself from intrinsically more resistant processes and materials — sapphire substrates, transistor geometries less sensitive to particle-induced charge — the classical rad-hard approach used in scientific and institutional missions, but one that typically relies on fabrication processes several generations behind the most recent commercial AI accelerators, at a cost in density and energy efficiency. The second strategy, the one Google, Nvidia and Starcloud are exploring, is to take commercial off-the-shelf (COTS) silicon and add system-level mitigation: error-correcting-code (ECC) memory capable of detecting and correcting bit flips, redundant verification of critical calculations through triple modular redundancy, scheduled periodic "scrubbing" of configuration memory in programmable circuits, and anomaly-detection algorithms able to isolate a faulty compute core without halting the entire mission. Google reported that its TPU v6e chips passed radiation testing equivalent to a five-year low-Earth-orbit mission without meaningful functional degradation — a result suggesting that some recent commercial accelerators tolerate radiation better than previously assumed, though the finding does not necessarily generalize to other chip generations or higher-energy orbits.

Thermal management. Dissipating heat without convection requires rethinking cooling systems from the ground up. Solutions under study include large deployable radiators, two-phase fluid loops that carry heat from the chip to the radiator, and a reconsideration of compute density per module: an accelerator that on Earth is cooled with forced air or pressurized liquid may, in orbit, require a radiative surface several times larger than its own physical footprint, placing a practical ceiling on how much compute power can be concentrated in a single satellite.

Software-level fault tolerance and system architecture. Because no physical mitigation eliminates the probability of failure entirely, the software orchestrating these workloads must assume that single-event effects will occur at some statistically predictable rate. This favors distributed, redundant computing architectures spread across multiple satellites in a constellation, so that the temporary or permanent loss of one node does not compromise the entire task — in contrast to the monolithic, high-density architecture typical of a terrestrial training cluster.

Communication and optical links. The usefulness of an orbital data center also depends on its ability to exchange data between satellites and with the ground. Inter-satellite laser links, which Google has tested in the lab at speeds above 1.6 terabits per second, are essential for a constellation to function as a coherent computing system rather than a set of isolated nodes, but they introduce their own requirements for pointing accuracy and mechanical stability in an environment of constant vibration and thermal variation.

4. Consequences for Applications and Intended Uses

The survival requirements described above are not an implementation detail: they directly determine which AI workloads are viable in orbit within this decade's horizon, and which will remain the province of terrestrial infrastructure.

Inference before training. Training large-scale foundation models requires tight, low-latency synchronization across thousands of accelerators over weeks or months of continuous operation — a regime particularly vulnerable to intermittent node failures and to inter-satellite bandwidth limits, however fast optical links may be compared with the cabling inside a terrestrial data center. Inference, by contrast, is a workload more tolerant of latency and of the occasional loss of a node, making it the natural use case for this first generation of platforms. Starcloud's own milestone — querying a model already trained on Earth — illustrates this ordering of priorities.

Geospatial processing at the orbital edge. One of the applications with the strongest near-term economic case is onboard processing of Earth-observation data: satellite imagery, synthetic-aperture radar and other forms of remote sensing generate data volumes that today are transmitted raw to the ground for analysis. Running inference directly on the satellite — for instance, to detect changes, classify objects or filter out cloud cover before transmission — dramatically reduces the bandwidth required and shortens the time between image capture and the availability of usable information, which matters for agricultural monitoring, disaster response and geospatial-intelligence applications. Starcloud has already processed radar data from Capella Space's satellites under this scheme.

Constellation autonomy and space operations. A second field of application is autonomous decision-making within a satellite constellation: maneuver planning, power management, debris-conjunction detection and coordination among satellites without relying on a ground station for every decision. Nvidia has explicitly named this use case as one of the goals of its space-computing platforms.

Lifespan limits and replacement economics. The same accumulated-radiation degradation mechanisms that require mitigation also impose a finite and relatively short useful life — on the order of a few years — on hardware deployed in orbit, in a context where replacing or upgrading a faulty component cannot be solved by sending a technician, as it can in a terrestrial data center. This makes launch cost, satellite mass-production capacity and the planning of full-constellation renewal cycles just as decisive for the project's economic viability as the chip's own performance. Industry observers, including analysts skeptical of the current investor enthusiasm, have noted that the relevant comparison is not the cost per FLOP of an orbital chip versus a terrestrial one, but the total lifecycle cost of an entire constellation versus that of a dedicated power plant serving an equivalent terrestrial data center.

5. Discussion and Outlook

The evidence available in 2026 — Google's TPU radiation testing, the successful operation of a language model on a commercial GPU in orbit, and Nvidia's announcement of dedicated space-computing platforms — suggests that the physical survival of commercial silicon in low Earth orbit is a tractable near-term engineering problem rather than an insurmountable barrier. The real bottleneck appears to lie not simply in whether a chip can survive the radiation of a multi-year mission, but in whether the system architecture built around that chip — redundancy, optical communication, thermal management and fault tolerance — can sustain AI workloads at a scale and cost per unit of compute competitive with the terrestrial alternative, even accounting for the latter's energy and regulatory constraints.

It is therefore reasonable to expect that the next phase of this nascent industry will concentrate on distributed inference and Earth-observation data processing, where tolerance for latency and partial failure is greater, before it becomes economically defensible to move the training of the largest foundation models into orbit. The discipline that the space environment demands — designing for likely failure rather than ideal operation — may, in turn, leave behind engineering lessons of independent value for building more resilient AI infrastructure on Earth.

 

Glossary

●      Low Earth Orbit (LEO) — The region of space roughly 160–2,000 km above Earth's surface, where most current and proposed orbital data-center satellites operate.

●      Sun-synchronous orbit — A near-polar orbit timed so a satellite passes over the same location at the same local solar time each day, useful for maximizing continuous sunlight exposure for solar power.

●      Rad-hard (radiation-hardened) — Electronics manufactured with materials and design techniques specifically intended to resist damage from ionizing radiation, typically at the cost of raw performance and manufacturing recency.

●      COTS (commercial off-the-shelf) — Hardware, such as a standard Nvidia GPU or Google TPU, built for ordinary commercial use rather than designed from scratch for the space environment.

●      Single-event effect (SEE) — A malfunction caused by a single high-energy particle striking a circuit, ranging from a harmless bit flip to a destructive short circuit.

●      Single-event upset (SEU) — A single-event effect in which a particle strike flips the stored value of a memory bit without permanently damaging the hardware.

●      Latch-up — A potentially destructive short-circuit condition inside a chip, sometimes triggered by a single-event effect, that can permanently disable the device if not detected and interrupted quickly.

●      Total ionizing dose (TID) — Cumulative damage to a device's electrical properties caused by continuous, long-term exposure to radiation over the life of a mission, as distinct from single discrete particle strikes.

●      ECC (error-correcting code) memory — Memory that stores extra bits alongside data, allowing the system to detect and, in many cases, automatically correct bit errors such as those caused by radiation.

●      Triple modular redundancy (TMR) — A fault-tolerance technique in which a calculation is performed three times in parallel and the result decided by majority vote, so a single corrupted result is automatically outvoted.

●      Scrubbing — The practice of periodically rewriting or re-verifying a chip's configuration memory to correct any radiation-induced errors before they accumulate or cause a failure.

●      Outgassing — The release of trapped gases from materials such as adhesives or polymers when exposed to the vacuum of space, which can contaminate nearby optical or electronic surfaces.

●      Thermal cycling — Repeated heating and cooling of a spacecraft's components as it moves in and out of sunlight, causing mechanical fatigue distinct from the stable temperatures of a terrestrial server room.

●      Radiative cooling — Heat dissipation via infrared radiation into space, the only cooling mechanism available in vacuum, as opposed to the convective (air- or liquid-based) cooling used on Earth.

●      Micrometeoroid — A tiny natural particle, often smaller than a grain of sand, that can still cause damage on impact due to extremely high relative orbital velocities.

●      Orbital debris — Human-made fragments, such as pieces of defunct satellites or spent rocket stages, that remain in orbit and pose a collision risk to active spacecraft.

●      Inter-satellite optical (laser) link — A communication channel using laser light rather than radio waves to transmit data directly between satellites, enabling much higher bandwidth.

●      TPU (Tensor Processing Unit) — A custom AI accelerator chip designed by Google, distinct from general-purpose GPUs, optimized for machine-learning workloads.

 

 

References

●      NVIDIA Newsroom (2026). "NVIDIA Launches Space Computing, Rocketing AI Into Orbit." https://nvidianews.nvidia.com/news/space-computing

●      TechRepublic (March 18, 2026). "Nvidia Launches Space-Ready AI Platforms for Orbital Data Centers." https://www.techrepublic.com/article/news-nvidia-space-ai-chips-orbital-data-centers/

●      Data Center Dynamics (July 27, 2026). "Starcloud runs AI model in space." https://www.datacenterdynamics.com/en/news/starcloud-runs-ai-model-in-space/

●      Gizmodo (September 2026). "Google's Project Suncatcher Is Sending AI Chips Into Space Next Week." https://gizmodo.com/googles-project-suncatcher-is-sending-ai-chips-into-space-next-week-2000816985

●      Introl Blog (February 21, 2026). "Orbital Data Center Race 2026." https://introl.com/blog/orbital-data-centers-space-computing-race-2026

●      InvestorPlace (September 26, 2026). "5 Space Stocks to Buy as Google Takes AI Into Orbit With Project Suncatcher." https://investorplace.com/hypergrowthinvesting/2026/09/spacex-just-put-a-launch-date-on-the-orbital-ai-boom/

●      Tech Insider (September 2026). "Google Project Suncatcher: 4 TPUs Launch to Orbit Oct 1." https://tech-insider.org/google-project-suncatcher-orbital-ai-data-center-2026/