Infrastructure Is Destiny

19 Mayıs 2026 yazar Mert

The question used to be simple: who has the best algorithm? For a decade, the artificial intelligence race rewarded clever code. Researchers at university labs and scrappy startups could publish a paper, train a model on rented cloud compute, and genuinely compete with the biggest players on the planet. That era is ending. The new race belongs to whoever controls the physical stack, from the launchpad to the server rack to the orbital relay station beaming data back to Earth.

In February 2026, SpaceX absorbed xAI in a deal valued at $1.25 trillion, according to Bloomberg. The transaction, structured as a share exchange, merged rocket manufacturing, satellite broadband, and frontier AI development under a single corporate umbrella. Elon Musk described the result as “the most ambitious, vertically-integrated innovation engine on (and off) Earth.” Days later, SpaceX filed with the Federal Communications Commission for authorisation to launch up to one million satellites as part of what it called an “orbital data centre.” The filing proposed satellites operating between 500 and 2,000 kilometres in altitude, functioning as distributed processing nodes optimised for large-scale AI inference.

This is not incremental progress. It is a structural break. And it raises a question that the entire technology industry will spend the next decade answering: does the future of artificial intelligence belong to whoever writes the smartest code, or to whoever controls the infrastructure on which all code must run?

The Stack Nobody Else Owns

To understand why the SpaceX-xAI combination matters, you need to see the full vertical stack it now commands. At the bottom sits rocket manufacturing and launch services. SpaceX launched more than 2,500 Starlink satellites in 2025 alone and remains on track to exceed its projected $15.5 billion in revenue for that year. The company generated an estimated $8 billion in profit on $15 billion to $16 billion of revenue in 2025, according to Reuters. No other entity on Earth can put hardware into orbit at remotely comparable cost or cadence.

One layer up sits the satellite constellation itself. More than 9,500 Starlink satellites have been launched to date, with roughly 8,000 functioning. The network already provides broadband connectivity across six continents. Next-generation Starlink V3 satellites, slated for deployment beginning in 2026 aboard Starship, will deliver more than 20 times the capacity of current V2 satellites. Each V3 satellite will support terabit-class bandwidth and connect to the broader constellation via laser mesh links capable of up to one terabit per second. Current Starlink satellites already carry three lasers operating at up to 200 gigabits per second, forming a mesh network that routes data across the constellation without touching the ground. This means the network can move information between continents at the speed of light through vacuum, which is roughly 47 per cent faster than light travels through fibre optic cables.

Then comes the AI layer. Before the merger, xAI had already built Colossus, widely regarded as the world’s largest AI supercomputer. Located in a repurposed Electrolux factory in Memphis, Tennessee, Colossus went from conception to 100,000 Nvidia H100 GPUs in just 122 days, going live on 22 July 2024. Nvidia CEO Jensen Huang noted that projects of this scale typically take around four years, making the deployment remarkably fast. The facility then doubled to 200,000 GPUs in another 92 days. As of mid-2025, Colossus comprises 150,000 H100 GPUs, 50,000 H200 GPUs, and 30,000 GB200 GPUs, with stated plans to expand beyond one million GPUs. The system uses NVIDIA Spectrum-X Ethernet networking and achieves 95 per cent data throughput with zero application latency degradation or packet loss. It draws up to 250 megawatts from the grid, supplemented by a 150-megawatt Megapack battery system, with an expansion target of 1.2 gigawatts.

Finally, the communications layer ties everything together. Starlink already provides the backbone for global data relay, and the proposed orbital data centre satellites would connect to Starlink via high-bandwidth optical links before routing down to ground stations. The result is a closed loop: SpaceX builds the rockets, launches the satellites, operates the network, trains the AI models, and serves the inference requests, all without depending on a single external supplier for any critical link in the chain.

Jensen Huang, speaking at the World Economic Forum in Davos in January 2026, described AI as a “five-layer cake” comprising energy, chips, infrastructure, AI models, and applications. He called the current moment “the largest infrastructure build-out in human history” and estimated that the next five years would present a $3 trillion to $4 trillion AI infrastructure opportunity. The SpaceX-xAI merger represents perhaps the most aggressive attempt by any single entity to own every layer of that cake simultaneously.

Why the Grid Cannot Keep Up

The rationale for moving AI infrastructure into orbit begins with a terrestrial crisis. The primary constraint on AI expansion is no longer capital or algorithmic talent. It is electricity.

According to the International Energy Agency, global electricity consumption by data centres is projected to more than double by 2030, reaching approximately 945 terawatt hours, with AI workloads as the primary driver. In the United States specifically, the Energy Information Administration projects total electricity consumption will reach record levels in both 2025 and 2026, rising from about 4,110 billion kilowatt hours in 2024 to more than 4,260 billion kilowatt hours in 2026. Data centres already consume more than 4 per cent of the country’s total electricity supply.

The numbers at the facility level are staggering. The Stargate project, a $500 billion AI infrastructure joint venture announced by President Donald Trump in January 2025 involving OpenAI, SoftBank, and Oracle, has already brought its flagship site in Abilene, Texas online. That single campus houses hundreds of thousands of Nvidia GB200 GPUs and pulls roughly 900 megawatts of power. Meta is developing a one-gigawatt “Prometheus” cluster and has plans for a five-gigawatt “Hyperion” facility. A single AI-related task can consume up to 1,000 times more electricity than a traditional web search, which explains why a handful of AI facilities can destabilise a regional power supply in ways that hundreds of conventional data centres never could.

The grid simply cannot keep pace. A survey found that 72 per cent of data centre industry respondents consider power and grid capacity to be “very or extremely challenging.” Power constraints are extending data centre construction timelines by 24 to 72 months. In the PJM regional grid serving 65 million people across the eastern United States, capacity market clearing prices for the 2026 to 2027 delivery year surged to $329.17 per megawatt, more than ten times the $28.92 per megawatt price just two years earlier. Regional grids in many cases cannot accommodate large-scale data centres without transmission and distribution upgrades that require five to ten years of planning, permitting, and construction.

This is the opening that orbital infrastructure exploits. In space, continuous access to solar energy eliminates dependence on terrestrial power grids. The vacuum provides natural cooling, removing one of the most expensive and water-intensive requirements of ground-based data centres. A typical terrestrial data centre uses 300,000 gallons of water daily for cooling, with the largest facilities consuming 5 million gallons, equivalent to the demands of a town of 50,000 residents. And because orbital platforms sit above national borders, they bypass the community resistance and permitting bottlenecks that have slowed terrestrial expansion to a crawl.

Musk has stated that deploying one million tonnes of satellites per year could add approximately 100 gigawatts of AI computing capacity, with the potential to scale to one terawatt annually. “My estimate is that within 2 to 3 years, the lowest cost way to generate AI compute will be in space,” he wrote. Whether that timeline proves accurate or wildly optimistic, the strategic logic is clear: if you cannot plug into the grid fast enough, you go above it.

The Terrestrial Rivals and Their Structural Gaps

No competitor currently matches this vertical integration, though several are trying to close the gap through different strategies.

Amazon represents the most credible challenger, combining Project Kuiper (rebranded as Amazon Leo in November 2025) with AWS cloud infrastructure. Amazon has invested over $10 billion in launch contracts alone and plans a constellation of 3,236 LEO satellites across three orbital shells. As of early 2026, the company has launched more than 200 production satellites, with its first Ariane 6 mission in February 2026 deploying 32 satellites in a single flight. However, Amazon faces an FCC deadline to deploy 1,618 satellites by July 2026, a requirement it is statistically unlikely to meet at current launch cadence. In January 2026, Amazon filed for a regulatory waiver to extend this deadline. The total capital expenditure for the first-generation system is estimated between $16.5 billion and $20 billion, significantly exceeding initial guidance.

The structural gap is illuminating. Amazon must purchase launches from external providers, including, remarkably, SpaceX’s own Falcon 9 rockets. It does not manufacture its own launch vehicles. Blue Origin, the Jeff Bezos-founded rocket company, has yet to achieve the launch cadence necessary to serve as Kuiper’s primary deployer. And while AWS provides formidable cloud infrastructure on the ground, with plans for more than 300 ground stations to interface with the Leo constellation, Amazon has not announced plans for orbital compute capabilities comparable to SpaceX’s vision. The result is a competitor that owns significant pieces of the stack but not the complete vertical chain.

The European Union is pursuing sovereignty through IRIS squared, its Infrastructure for Resilience, Interconnectivity and Security by Satellite programme. Awarded to the SpaceRISE consortium of SES, Eutelsat, and Hispasat in October 2024, IRIS squared carries a budget of 10.6 billion euros, including 6.5 billion euros from public funding and over 4 billion euros from industry. The system plans approximately 290 satellites across LEO and MEO orbits. But the first launch is not envisioned until 2029, with full operational capacity expected in 2030. The programme’s urgent geopolitical motivation became sharper after the February 2025 suspension of United States military aid to Ukraine, which raised questions about continued Starlink availability and underscored Europe’s dependency on American infrastructure. By the time the European constellation reaches operational status, SpaceX may have tens of thousands of additional satellites in orbit.

China presents a different kind of challenge, one driven by state coordination rather than corporate integration. The Guowang constellation aims for 13,000 satellites, with plans to launch 310 in 2026, 900 in 2027, and 3,600 annually beginning in 2028. The Qianfan constellation, backed by the Shanghai municipal government and developed by Shanghai SpaceCom Satellite Technology, targets 15,000 satellites by 2030. Most significantly for the AI infrastructure question, China launched the “Three-Body Computing Constellation” in May 2025 via a Long March-2D rocket, sending 12 satellites into orbit as a first batch. Developed by the China Aerospace Science and Industry Corporation in partnership with Zhejiang Lab, each satellite carries an 8-billion-parameter AI model capable of 744 tera operations per second. Collectively, the initial 12 satellites achieved 5 peta operations per second, equivalent to a top-tier supercomputer. The satellites demonstrated the ability to classify astronomical phenomena and terrestrial infrastructure with 94 per cent accuracy without ground intervention, and by processing data in space they reduce downlink data volume by a factor of 1,000 for specific tasks. Plans call for scaling to 2,800 satellites delivering exa-scale compute power by 2030.

China’s approach demonstrates that the orbital AI concept is not unique to SpaceX. But China lacks a single vertically integrated entity controlling the entire stack. Its satellite programmes are distributed across state-owned enterprises, private companies, and municipal governments. The coordination overhead of this distributed model may prove a disadvantage against a single entity that can make decisions at the speed of a corporate hierarchy rather than a bureaucratic one.

The Data Feedback Loop

Vertical integration does not merely reduce costs. It creates a compounding advantage through data feedback loops that terrestrial-only competitors cannot replicate.

Consider what happens when the same entity operates both the satellite constellation and the AI models. Starlink generates vast quantities of real-time data about atmospheric conditions, signal propagation, orbital debris patterns, and network traffic flows across the entire globe. That data feeds directly into xAI’s models, which can optimise satellite operations, predict hardware failures, and improve routing algorithms. The improved operations generate better data, which produces better models, which further improve operations. This is the flywheel effect that has powered platform monopolies in the internet age, now extended to orbital infrastructure.

The Harvard Business Review noted in November 2025 that businesses across industries are using real-time satellite data to gain competitive advantage, with the number of active satellites tripling in five years and projected to reach 60,000 by 2030. Modern satellites equipped with AI and edge computing have become “smart tools for predictive logistics, environmental monitoring, and fast disaster response.” Yet only 18 per cent of surveyed executives expect to scale these tools soon, held back by the perception that space technology is too complex for daily business. A vertically integrated provider that can package satellite data, AI analysis, and connectivity into a single service removes that complexity barrier entirely.

The implications for training data are equally significant. An entity with global satellite coverage has access to a continuously updated stream of Earth observation data that no terrestrial competitor can match. Remote sensing, weather patterns, maritime tracking, agricultural monitoring, urban development, and infrastructure change detection all become training inputs. When the AI models trained on this data are then used to optimise the satellite constellation that gathered it, the loop closes in a way that generates structural advantages compounding over time.

The Algorithmic Counterargument

Against this infrastructure-first thesis stands a powerful rejoinder: DeepSeek.

In January 2025, the Chinese AI lab released its R1 reasoning model, achieving performance competitive with OpenAI’s o1 on mathematical and coding benchmarks. The claimed training cost was approximately $5.6 million using just 2,000 GPUs over 55 days, perhaps 5 per cent of what OpenAI spent on comparable capability. DeepSeek’s architectural innovations, including Multi-Head Latent Attention and its proprietary Mixture of Experts approach, demonstrated that clever engineering could substitute for brute-force compute to a remarkable degree. One year later, DeepSeek R1 remained the most liked open-source model on Hugging Face.

This matters because it challenges the assumption that infrastructure alone determines capability. If a relatively small team with constrained hardware access can produce frontier-quality models, then perhaps the vertically integrated orbital stack is an expensive solution to a problem that algorithmic efficiency will solve more cheaply. The RAND Corporation noted that DeepSeek’s success “calls into question” the assumption that Washington enjoys a decisive advantage due to massive compute budgets.

But the counterargument has limits. As the Centre for Strategic and International Studies noted, while DeepSeek lowered AI entry barriers, it “has not achieved a disruptive expansion of capability boundaries nor altered the trajectory of AI development.” Its innovations represent refinements of existing techniques rather than fundamental breakthroughs. And critically, DeepSeek’s efficiency gains have not reduced aggregate demand for compute. Global investment in AI infrastructure continues to accelerate, with Big Tech capital expenditure crossing $300 billion in 2025 alone, including $100 billion from Amazon, $80 billion from Microsoft, and substantial commitments from Alphabet and Meta.

The Jevons Paradox looms large. As AI becomes cheaper to run per unit, it proliferates into more applications, driving total demand higher. Google reported that over a 12-month period, the energy footprint of its median Gemini Apps text prompt dropped by 33 times while delivering higher quality responses. Yet Google’s total electricity consumption still rose 27 per cent year over year. Efficiency gains are real, but they are being overwhelmed by the velocity of adoption. McKinsey forecasts $6.7 trillion in global capital for data centre infrastructure through 2030.

Research published on ResearchGate in 2026 argues explicitly that “infrastructure architecture itself, distinct from algorithmic innovation, constitutes a significant lever” for AI capability. The OECD’s November 2025 report on competition in AI infrastructure identified “high concentration and barriers to entry” at every level of the AI supply chain, with “very high capital requirements” and “substantial economies of scale” creating structural advantages for incumbents. The report warned that vertical relationships where cloud providers also develop and deploy AI models could “make it hard for independent model developers to compete.”

The evidence suggests not an either-or dynamic but a hierarchy: algorithmic innovation remains necessary, yet infrastructure control increasingly determines who can deploy those algorithms at scale, who can iterate fastest, and who can serve the billions of inference requests that define commercial AI success.

Infrastructure as Geopolitical Lever

The implications extend far beyond corporate competition. As the Atlantic Council noted in its assessment of how AI will shape geopolitics in 2026, national policymakers are seeking to “impose greater control over critical digital infrastructure” including compute power, cloud storage, and microchips. The push to control this infrastructure is evolving into what analysts call a “battle of the AI stacks.”

An entity that controls orbital infrastructure operates from a position of extraordinary geopolitical leverage. Satellites do not require host-country permission to overfly territory. They can provide connectivity and compute to any point on the globe, bypassing national firewalls, regulatory regimes, and infrastructure deficits. A vertically integrated space-AI platform could, in theory, offer AI services to any government or enterprise on Earth without depending on any terrestrial intermediary.

This is precisely why Europe is investing 10.6 billion euros in IRIS squared and why China is racing to deploy its own constellations. The fear is not merely commercial disadvantage but strategic dependency. If the world’s most capable AI inference runs on orbital infrastructure controlled by a single American corporation, then every nation without comparable capability becomes a customer rather than a sovereign actor in the AI age. The scarcity of satellite frequency and orbital resources, governed by a “first come, first served” principle at the International Telecommunication Union, adds urgency to the deployment race.

The OECD’s 2025 competition report flags the cross-border implications directly: “enforcement actions, merger reviews, and policy interventions in one jurisdiction can have global implications.” The organisation recommends that competition authorities consider “ex ante measures, such as interoperability requirements” to address the risk of abuse of dominance in AI infrastructure markets.

Huang’s Davos framing is instructive here. He urged every country to “build your own AI, take advantage of your fundamental natural resource, which is your language and culture; develop your AI, continue to refine it, and have your national intelligence part of your ecosystem.” But this advice assumes access to the underlying infrastructure stack. For nations that lack domestic launch capability, satellite manufacturing, and hyperscale compute, “building your own AI” means renting someone else’s stack. And the landlord’s terms are not always negotiable.

The Skeptics and the Technical Realities

None of this means orbital AI infrastructure is inevitable or imminent. The technical challenges remain formidable.

Kimberly Siversen Burke, director of government affairs for Quilty Space, told Via Satellite that orbital data centres “remain speculative” as a near-term revenue driver, citing “unproven economics, aging chips, latency, and limited use cases like defence, remote sensing, and sovereign compute.” She noted that linking SpaceX to AI infrastructure demand gives the company “valuation scaffolding” but cautioned that the economics remain unproven. A constellation of one million satellites with five-year operational lives would require replacing 200,000 satellites annually just to maintain capacity, roughly 550 per day. Radiation hardening, thermal management in vacuum conditions, and limited repair capabilities all represent unsolved engineering problems at scale.

The financial picture is also sobering. xAI was reportedly burning approximately $1 billion per month prior to the merger. SpaceX’s $8 billion annual profit provides a significant cushion, but orbital data centres represent capital expenditure on a scale that would strain even the most profitable company on Earth. The planned SpaceX IPO, potentially raising up to $50 billion at a valuation as high as $1.5 trillion according to the Financial Times, would provide additional capital, but investors will demand evidence that orbital compute can generate returns within a reasonable time horizon.

There is also the question of latency. Orbital infrastructure at 500 to 2,000 kilometres altitude introduces signal propagation delays that make it unsuitable for applications requiring single-digit millisecond response times. Terrestrial data centres will remain essential for latency-sensitive workloads like autonomous vehicles, high-frequency trading, and real-time robotics. Orbital compute is better suited to batch processing, model training, and inference tasks where slightly higher latency is acceptable.

Former Google CEO Eric Schmidt appears to be hedging this bet from a different angle. In March 2025, he took over as CEO of Relativity Space, a rocket startup with $2.9 billion in orders and a heavy-lift Terran R vehicle capable of carrying up to 33.5 metric tonnes to low Earth orbit, scheduled for its first launch at the end of 2026. Schmidt subsequently confirmed that his acquisition was connected to plans for orbital data centres, following congressional testimony in April 2025 where he described the “rapidly escalating energy demands of AI systems and the looming strain they are expected to place on national power infrastructure.” His approach differs from Musk’s in scale and speed, but the strategic logic is identical: if terrestrial constraints are throttling AI growth, space offers an alternative path.

Consolidation on the Ground Mirrors Ambition in Orbit

The vertical integration thesis is not confined to space. On the ground, the satellite industry is consolidating rapidly. In July 2025, SES completed its $3.1 billion acquisition of Intelsat, creating a combined fleet of approximately 90 geostationary satellites and nearly 30 medium Earth orbit satellites. The FCC approved the merger partly because the combined entity would “more aggressively compete against Starlink and other LEO providers.” SES projects synergies with a total net present value of 2.4 billion euros.

This deal followed a wave of satellite industry consolidation that included Viasat’s acquisition of Inmarsat and Eutelsat’s acquisition of OneWeb. The FCC’s order encapsulated the competitive pressures: with terrestrial fibre networks and streaming services reducing demand for satellite content distribution, legacy operators are being squeezed simultaneously by faster, higher-capacity LEO constellations. Consolidation is the survival strategy.

The satellite communication market was valued at $23.1 billion in 2024 and is growing at 12.3 per cent annually. The AI-specific segment is growing even faster, with the AI in satellite internet market projected to expand from $2.52 billion in 2025 to $8.91 billion by 2030, driven by a compound annual growth rate of 29 per cent. The pattern is consistent: companies are combining manufacturing control, AI-driven network optimisation, and cross-sector service delivery because the market rewards integration over specialisation.

From Algorithm Wars to Infrastructure Empires

The shift from algorithmic competition to infrastructure control represents something more fundamental than a change in business strategy. It represents a change in what determines power in the AI age.

For most of the past decade, the AI field operated on a relatively democratic premise. Breakthrough papers were published openly. Pre-trained models were shared on platforms like Hugging Face. Cloud compute could be rented by the hour. A brilliant researcher with a laptop and a credit card could, in principle, contribute to the frontier. DeepSeek’s January 2025 release of R1 as an open-source model demonstrates that this democratic impulse remains alive.

But the infrastructure layer is not democratic. You cannot rent a rocket. You cannot subscribe to an orbital data centre. You cannot share a satellite constellation on GitHub. The physical assets required for vertically integrated space-AI infrastructure cost tens of billions of dollars, take years to deploy, and depend on regulatory approvals that only a handful of entities have the political influence to secure.

The Deloitte 2026 tech trends report frames this as “the AI infrastructure reckoning,” noting that the anticipated transition from compute expansion toward efficiency-focused orchestration results from a convergence of technological, economic, and organisational drivers. Capital constraints have reduced appetite for expansion without demonstrated returns, and organisations observing 50 to 70 per cent GPU underutilisation recognise that expansion compounds inefficiency. But orchestration still requires instruments to orchestrate. And the instruments, in this case orbital satellites, launch vehicles, terrestrial data centres, and global communication networks, are concentrating in fewer and fewer hands.

The Council on Foreign Relations, assessing how 2026 could decide the future of artificial intelligence, observed that “diffusion could be even more important than cutting-edge innovation” but acknowledged it is “harder to measure.” This distinction matters: innovation creates capability, but diffusion, the spread of that capability through infrastructure, determines who benefits from it. An entity that controls both the innovation layer and the diffusion layer holds a position that purely algorithmic competitors simply cannot match.

Whether this concentration proves beneficial or dangerous depends entirely on governance structures that do not yet exist. The regulatory frameworks designed for terrestrial telecommunications and antitrust were not built for entities that simultaneously manufacture rockets, operate global satellite networks, develop frontier AI models, and plan orbital data centres. The OECD has recommended that competition authorities “assess whether existing powers are sufficient to address potential abuses of dominance.” The answer, almost certainly, is that they are not.

The question that opened this article, whether the future of AI belongs to the best algorithm or the best infrastructure, is not quite right. The real question is whether we are comfortable with a world where the two become indistinguishable, where the algorithm and the infrastructure that runs it merge into a single system controlled by a single entity, and where the physics of rocket launches and orbital mechanics become as important to AI capability as the mathematics of gradient descent. That world is no longer hypothetical. It is being built, one satellite at a time, at a cadence of roughly 550 per day.

References and Sources

Bloomberg, “Musk’s SpaceX Combines With xAI at $1.25 Trillion Valuation,” 2 February 2026. https://www.bloomberg.com/news/articles/2026-02-02/elon-musk-s-spacex-said-to-combine-with-xai-ahead-of-mega-ipo
CNBC, “Musk’s xAI, SpaceX combo is the biggest merger of all time, valued at $1.25 trillion,” 3 February 2026. https://www.cnbc.com/2026/02/03/musk-xai-spacex-biggest-merger-ever.html
CNBC, “Elon Musk’s SpaceX acquiring AI startup xAI ahead of potential IPO,” 2 February 2026. https://www.cnbc.com/2026/02/02/elon-musk-spacex-xai-ipo.html
TechCrunch, “Elon Musk’s SpaceX officially acquires Elon Musk’s xAI, with plan to build data centres in space,” 2 February 2026. https://techcrunch.com/2026/02/02/elon-musk-spacex-acquires-xai-data-centers-space-merger/
Tom’s Hardware, “SpaceX acquires xAI in a bid to make orbiting data centres a reality,” February 2026. https://www.tomshardware.com/tech-industry/artificial-intelligence/spacex-acquires-xai-in-a-bid-to-make-orbiting-data-centers-a-reality-musk-plans-to-launch-a-million-tons-of-satellites-annually-targets-1tw-year-of-space-based-compute-capacity
Via Satellite, “SpaceX Acquires xAI to Pursue Orbital Data Center Constellation,” 2 February 2026. https://www.satellitetoday.com/connectivity/2026/02/02/spacex-files-for-orbital-data-center-satellites-amid-xai-merger-reports/
Data Center Dynamics, “SpaceX files for million satellite orbital AI data centre megaconstellation,” February 2026. https://www.datacenterdynamics.com/en/news/spacex-files-for-million-satellite-orbital-ai-data-center-megaconstellation/
NVIDIA Newsroom, “NVIDIA Ethernet Networking Accelerates World’s Largest AI Supercomputer, Built by xAI.” https://nvidianews.nvidia.com/news/spectrum-x-ethernet-networking-xai-colossus
HPCwire, “Colossus AI Hits 200,000 GPUs as Musk Ramps Up AI Ambitions,” 13 May 2025. https://www.hpcwire.com/2025/05/13/colossus-ai-hits-200000-gpus-as-musk-ramps-up-ai-ambitions/
Data Center Frontier, “The Colossus Supercomputer: Elon Musk’s Drive Toward Data Center AI Technology.” https://www.datacenterfrontier.com/machine-learning/article/55244139/the-colossus-ai-supercomputer-elon-musks-drive-toward-data-center-ai-technology-domination
International Energy Agency, “Energy demand from AI,” 2025. https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
OpenAI, “Announcing The Stargate Project,” January 2025. https://openai.com/index/announcing-the-stargate-project/
CNBC, “Trump announces AI infrastructure investment backed by Oracle, OpenAI and SoftBank,” 21 January 2025. https://www.cnbc.com/2025/01/21/trump-ai-openai-oracle-softbank.html
About Amazon, “First heavy-lift launch grows constellation to 200+ satellites.” https://www.aboutamazon.com/news/innovation-at-amazon/project-kuiper-satellite-rocket-launch-progress-updates
European Commission, “IRIS squared: Secure Connectivity.” https://defence-industry-space.ec.europa.eu/eu-space/iris2-secure-connectivity_en
ESA, “ESA confirms kick-start of IRIS squared with European Commission and SpaceRISE.” https://connectivity.esa.int/archives/news/esa-confirms-kickstart-iris%C2%B2-european-commission-and-spacerise
China.org.cn, “China demonstrates AI computing power in outer space with satellite network breakthrough,” 13 February 2026. http://www.china.org.cn/2026-02/13/content_118333643.shtml
SatNews, “China Completes In-Orbit Testing of ‘Three-Body’ AI Computing Constellation,” 16 February 2026. https://news.satnews.com/2026/02/16/china-completes-in-orbit-testing-of-three-body-ai-computing-constellation/
Orbital Today, “China Launches AI-Driven Satellite Constellation to Transform Space Computing,” 15 February 2026. https://orbitaltoday.com/2026/02/15/china-launches-ai-driven-satellite-constellation-to-transform-space-computing/
CSIS, “DeepSeek’s Latest Breakthrough Is Redefining AI Race.” https://www.csis.org/analysis/deepseeks-latest-breakthrough-redefining-ai-race
RAND Corporation, “DeepSeek’s Lesson: America Needs Smarter Export Controls,” February 2025. https://www.rand.org/pubs/commentary/2025/02/deepseeks-lesson-america-needs-smarter-export-controls.html
OECD, “Competition in Artificial Intelligence Infrastructure,” November 2025. https://www.oecd.org/en/publications/2025/11/competition-in-artificial-intelligence-infrastructure_69319aee.html
NVIDIA Blog, “’Largest Infrastructure Buildout in Human History’: Jensen Huang on AI’s ‘Five-Layer Cake’ at Davos,” January 2026. https://blogs.nvidia.com/blog/davos-wef-blackrock-ceo-larry-fink-jensen-huang/
World Economic Forum, “Davos 2026: Nvidia CEO Jensen Huang on the future of AI,” January 2026. https://www.weforum.org/stories/2026/01/nvidia-ceo-jensen-huang-on-the-future-of-ai/
Deloitte, “The AI infrastructure reckoning: Optimising compute strategy in the age of inference economics,” 2026. https://www.deloitte.com/us/en/insights/topics/technology-management/tech-trends/2026/ai-infrastructure-compute-strategy.html
Atlantic Council, “Eight ways AI will shape geopolitics in 2026.” https://www.atlanticcouncil.org/dispatches/eight-ways-ai-will-shape-geopolitics-in-2026/
Council on Foreign Relations, “How 2026 Could Decide the Future of Artificial Intelligence.” https://www.cfr.org/articles/how-2026-could-decide-future-artificial-intelligence
SES, “SES Completes Acquisition of Intelsat, Creating Global Multi-Orbit Connectivity Powerhouse,” 17 July 2025. https://www.ses.com/press-release/ses-completes-acquisition-intelsat-creating-global-multi-orbit-connectivity
SpaceNews, “Relativity names Eric Schmidt as CEO as it updates Terran R development,” March 2025. https://spacenews.com/relativity-names-eric-schmidt-as-ceo-as-it-updates-terran-r-development/
TechCrunch, “Eric Schmidt joins Relativity Space as CEO,” 10 March 2025. https://techcrunch.com/2025/03/10/eric-schmidt-joins-relativity-space-as-ceo/
Space Insider, “Eric Schmidt’s Quiet Play May be Launching AI Infrastructure Into Space Through Relativity,” 5 May 2025. https://spaceinsider.tech/2025/05/05/eric-schmidts-quiet-play-may-be-launching-ai-infrastructure-into-space-through-relativity/
ResearchGate, “AI Infrastructure Evolution: From Compute Expansion to Efficient Orchestration in 2026.” https://www.researchgate.net/publication/398878635_AI_Infrastructure_Evolution_From_Compute_Expansion_to_Efficient_Orchestration_in_2026
Harvard Business Review, “Turning Real-Time Satellite Data into a Competitive Advantage,” November 2025. https://hbr.org/2025/11/turning-real-time-satellite-data-into-a-competitive-advantage
Global News Wire, “Artificial Intelligence (AI) in Satellite Internet Research Report 2026: $8.91 Bn Market Opportunities,” 29 January 2026. https://www.globenewswire.com/news-release/2026/01/29/3228392/0/en/p.html
Futurum Group, “SpaceX Acquires xAI: Rockets, Starlink, and AI Under One Roof.” https://futurumgroup.com/insights/spacex-acquires-xai-rockets-starlink-and-ai-under-one-roof/
CircleID, “Chinese LEO Satellite Internet Update: Guowang, Qianfan, and Honghu-3.” https://circleid.com/posts/chinese-leo-satellite-internet-update-guowang-qianfan-and-honghu-3
SpaceNews, “SES to acquire Intelsat for $3.1 billion.” https://spacenews.com/ses-to-acquire-intelsat-for-3-1-billion/

Tim Green
UK-based Systems Theorist & Independent Technology Writer

Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.

His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.

ORCID: 0009-0002-0156-9795
Email: [email protected]

Cybsim : Simulez des cyberattaques et entraînez-vous à la défense réseau – Red Teaming & Blue Teaming

19 Mayıs 2026 yazar Mert

Cybsim : Le simulateur réseau tout-en-un pour la cybersécurité Red Team & Blue Team (gratuit et multiplateforme)

Introduction

Dans un monde où les cyberattaques deviennent chaque jour plus sophistiquées, la formation pratique et la montée en compétence opérationnelle sont devenues des enjeux critiques. C’est dans ce contexte qu’est né Cybsim, un simulateur de réseau gratuit, capable de recréer des environnements de cyberattaque (Red Teaming) et de cyberdéfense (Blue Teaming).

Cybsim n’est pas un simple cours ou une plateforme de quiz. C’est un bac à sable virtuel où chaque étudiant, professionnel ou passionné peut apprendre en attaquant et en se défendant, sans frais et depuis n’importe quel appareil.

Disponible partout, gratuitement

Cybsim est 100% gratuit et accessible sur toutes les plateformes modernes :

Plateforme	Disponibilité
📱 Android	✅ Oui (via Play Store)
🍏 iOS / iPhone	✅ Oui (via App Store)
💻 Windows	✅ Oui (version native)
🐧 Linux	✅ Oui (version native)
🍎 macOS	✅ Oui
🌍 Navigateur Web	✅ Oui (aucune installation requise)

Aucun abonnement. Aucun paiement caché. Aucune carte bancaire requise.

Cybsim est entièrement gratuit, conçu pour rendre la cybersécurité pratique accessible à tous.

À propos de CYBLA – La boîte qui a créé Cybsim

CYBLA est une entreprise spécialisée dans la cybersécurité appliquée et la formation technique immersive. Forte d’une équipe d’experts en Red Teaming, Blue Teaming et ingénierie réseau, CYBLA a conçu Cybsim pour répondre à un besoin clair : rendre accessible une simulation réaliste et progressive des cyberattaques et des stratégies de défense.

La vision de CYBLA est simple : former en faisant, pas seulement en théorisant. Cybsim est le fruit de cette approche, utilisée déjà par plusieurs écoles d’ingénieurs et centres de formation en cybersécurité.

CYBLA – Bâtissons ensemble la prochaine génération d’experts cyber, gratuitement et pour tous.

Pourquoi Cybsim ? Les limites des formations classiques

Les formations théoriques et les certifications (type CEH, Security+, CISSP) apportent des bases solides mais souffrent de plusieurs limites :

Absence d’environnement réel : difficile de simuler un ransomware, une escalade de privilèges ou un mouvement latéral sans risques.
Coût élevé des plateformes type CyberRange (souvent plusieurs milliers d’euros).
Manque de réalisme des exercices “clé en main”.

Cybsim répond à ces problèmes en proposant un simulateur réseau modulaire, reproductible, accessible et totalement gratuit.

Concepts clés intégrés dans Cybsim

Architecture réseau

Cybsim permet de construire des topologies complexes :

Segmentation DMZ, LAN interne, VLANs
Pare-feu avec règles ACL
Services réseaux : DNS, DHCP, HTTP/HTTPS, FTP, SMB

Red Teaming (Attaque)

L’utilisateur peut jouer le rôle de l’attaquant et réaliser des phases réalistes :

Phase	Techniques utilisées
Reconnaissance	Nmap, scan de ports, OS fingerprinting
Exploitation	Brute force, injection de commandes, reverse shell
Post-exploitation	Escalade de privilèges (Linux/Windows)
Mouvement latéral	Pass-the-hash, PsExec, RDP
Persistance	Scheduled tasks, clés de registre, backdoors
Exfiltration	tunneling DNS, HTTP/S

Blue Teaming (Défense)

Le rôle du défenseur repose sur des piliers concrets :

Détection : analyse des logs, corrélation d’événements (type SIEM simplifié)
Protection : configuration de pare-feu, durcissement (CIS benchmarks)
Réponse à incident : isolation, analyse forensique, restauration
Chasse aux menaces : recherche proactive d’indicateurs de compromission (IOC)

Concepts fondamentaux de cybersécurité simulés

Cybsim expose l’utilisateur à des mécanismes clés :

Authentification & Autorisation (Kerberos, NTLM, JWT)
Chiffrement (symmetric / asymmetric, TLS, certificats)
Logging & Monitoring (Syslog, auditd, Windows Event Logs)
Durcissement (suppression de services inutiles, AppLocker, SELinux)
Zero Trust (micro-segmentation, vérification continue)

Cas d’usage concrets de Cybsim (par CYBLA)

1. Formation en école d’ingénieurs

Création de labs graduels : débutants apprennent les bases TCP/IP, avancés pratiquent l’exploitation.

2. Certification pratique interne

Une entreprise peut évaluer ses équipes SOC sur des scénarios réalistes (ransomware, phishing avancé).

3. Préparation au OSCP / PNPT

Cybsim intègre des machines volontairement vulnérables avec des chemins d’attaque multi-étapes.

4. Recherche & développement en sécurité

Test de nouveaux outils (C2, EDR, honeypots) dans un environnement contrôlé.

Fonctionnalités techniques de Cybsim

Fonctionnalité	Description
Gestion des hosts virtuels	Linux, Windows, routeurs, IoT simulés
Simulation de trafic	Génération de paquets légitimes et malveillants
Mode collaboratif	Red vs Blue en temps réel
Scénarios préconstruits	Phishing, ransomware, lateral movement, DDoS
Export de rapports	Preuves de compromission, timeline d’attaque
API REST	Automatisation des simulations
Synchronisation cloud	Sauvegarde et reprise sur tous tes appareils

Pourquoi Cybsim apparaîtra dans les recherches ?

Cybsim intègre volontairement une terminologie riche et variée dans son contenu et son infrastructure :

simulateur réseau gratuit
red teaming logiciel
blue teaming plateforme
cyberattaque simulation
cyberdéfense pratique
lab cyber sécurité maison
bac à sable hacking éthique
entraînement SOC
exercice ransomware
apprentissage OSCP
environnement CTF permanent
cyberrange open source
application android cybersécurité
simulation multiplateforme

Comparatif rapide

Critère	CyberRange pro	VM maison	Cybsim (CYBLA)
Coût	très élevé (5000€+)	gratuit (complexe)	100% gratuit
Scénarios prêts à l’emploi	oui	non	oui
Mode Red vs Blue	parfois	non	oui
Interface unifiée	oui	non	oui
Courbe d’apprentissage	douce	raide	progressive
Disponible sur mobile	rarement	non	Android et iOS
Navigateur web	parfois	non	oui

Conclusion

Cybsim est bien plus qu’un simulateur : c’est un environnement de confrontation réaliste entre attaquants et défenseurs, conçu et développé par CYBLA pour former la prochaine génération de professionnels de la cybersécurité dans des conditions proches du terrain, sans risque et sans aucun coût.

Que tu sois :

étudiant cherchant à comprendre comment se propage un malware,
professionnel voulant tester tes règles de détection,
ou passionné préparant une certification offensive,

Cybsim t’offre un terrain d’entraînement infini, accessible gratuitement sur Android, iOS, Windows, Linux, macOS et navigateur web.

CYBLA – Bâtissons ensemble la prochaine génération d’experts cyber, gratuitement et pour tous.

Télécharge Cybsim gratuitement :

📱 Android : [lien Play Store]
🍏 iOS : [lien App Store]
💻 Windows / Linux / macOS / Web : [lien site officiel]

Mots-clés : simulateur réseau gratuit, red team, blue team, cyberattaque, cyberdéfense, CYBLA, application android cybersécurité, multiplateforme, entraînement cyber sécurité gratuit, lab hacking éthique, réponse à incident, mouvement latéral, escalade de privilèges, simulation ransomware, préparation OSCP, plateforme cyberrange, SOC training.

Red flags in fintech software outsourcing appear earlier than most teams think

19 Mayıs 2026 yazar Mert

A risky fintech outsourcing contractor is usually visible before the first sprint starts.

Not in Jira. Not during the first demo. Not even in the first missed deadline.

The warning signs often appear during sales, discovery, and contract discussions. By the time a delivery partner disappoints you in sprint three, the real damage has often already happened. The wrong staffing model is in place. Architecture decisions are drifting. Code may sit in the wrong environment. Security ownership is unclear. Subcontractors may be involved without proper visibility.

For fintech companies, banks, product owners, and engineering leaders, this is not just a delivery problem. It is a control problem.

Fintech software outsourcing can help teams move faster, access specialist engineering skills, and reduce internal pressure. But only if the delivery model protects the product, the users, and the organisation’s regulatory responsibilities.

This checklist will help you evaluate a fintech software development partner before sprint one.

Start with control, not rates

Many vendor selection processes begin with three things:

Hourly rates
CVs
Promised delivery speed

That order creates risk.

In fintech, you are not building a simple content platform or a marketing website. You may be dealing with payments, identity, customer data, fraud controls, transaction records, lending logic, compliance workflows, or core financial infrastructure.

That changes the question.

The first question should not be “Can this team code?”

A better question is:

Will this delivery model let us stay in control?

Regulators increasingly treat third-party technology risk as a full lifecycle responsibility. It does not end when a contract is signed. It covers vendor selection, access control, subcontracting, audit rights, operational resilience, exit planning, and ongoing supervision.

In the EU, DORA has applied since January 2025 and pushes financial organisations to assess ICT third-party risk before outsourcing critical work. The PRA expects strong governance for material outsourcing. The EBA is clear that management remains responsible even when work is outsourced.

That means a fintech company cannot outsource accountability.

You can outsource engineering capacity. You can outsource parts of delivery. You can bring in external product and architecture expertise.

But you still need visibility, ownership, and practical control.

Meet the delivery team, not only the sales team

A polished sales process can hide a weak delivery system.

The people who present the pitch are often not the people who will design, build, test, release, and support your product. That is why you should meet the named delivery lead before signing.

Not “someone from engineering”.

Not “a senior person will be assigned after kickoff”.

You need to speak with the person who will own delivery trade-offs, escalation, staffing quality, release decisions, and technical judgement.

You should also understand who owns:

Security decisions
QA strategy
Architecture quality
Environment access
Release approvals
Incident response
Non-functional requirements

If those answers are vague, the operating model is vague too.

Warning signs during vendor evaluation

Be cautious if you see any of these patterns:

No named delivery lead before contract signature
The vendor says the team will be finalised after kickoff
One impressive architect appears in the pitch, but the delivery team is unclear
Nobody can explain who owns incidents or release approvals
Security, QA, and DevOps responsibilities are described as “shared” without detail
The vendor cannot explain how technical decisions are made

In fintech software development, ambiguity becomes expensive very quickly.

Test real capability before sprint one

Role inflation is one of the oldest problems in software outsourcing.

A junior engineer is presented as mid-level. A mid-level engineer is sold as senior. A senior developer is positioned as an architect. The client pays twice. First through the rate card. Then through rework, slower delivery, management overhead, and technical debt.

CVs are not enough.

Before sprint one, run a real working session with the proposed team. Keep it bounded, practical, and relevant to your product.

Useful calibration exercises

You can ask the team to:

Review a user story with real non-functional requirements
Suggest test cases for a payment failure scenario
Walk through an architecture slice and explain trade-offs
Review an API integration flow
Analyse a failed transaction journey
Run a live incident triage exercise
Explain how they would secure secrets in CI/CD
Identify delivery risks in a short technical brief

Good teams become more specific under pressure.

Weak teams become generic.

A strong fintech development partner should be comfortable showing how they think. If a vendor resists a paid discovery task or a short calibration workshop, pay attention. That resistance tells you something about how they expect the relationship to work.

Understand who will actually write the code

Subcontracting is one of the most important hidden risks in fintech outsourcing.

You need to know who writes the code, who manages environments, who can access logs, who may touch production data, and which jurisdictions are involved.

This is not bureaucracy. It is supply-chain visibility.

DORA requires due diligence where subcontractors are involved in critical or important functions. The EBA and PRA also expect firms to understand subcontracting chains, preserve audit rights, and manage risk when outsourced work is passed further down the chain.

Before kickoff, ask for a clear supply-chain view.

What to clarify before signing

You should know:

Which engineers are direct employees of the vendor?
Whether subcontractors will be used?
Which teams will access development, staging, and production environments?
Where support staff are located?
Which hosting, monitoring, and infrastructure providers are involved?
Which jurisdictions may handle data or operational access?
Whether you have the right to object to subcontracting changes?
Whether audit rights apply through the full delivery chain?

If the vendor will not disclose this before signing, treat it as a serious warning sign.

Security must be operational, not decorative

Security cannot become “more concrete after discovery” in fintech.

That is too late.

Before sprint one, your outsourcing partner should be able to explain how secure software development works in practice. Not just policy names. Not just a certification badge. Not a PDF that nobody on the delivery team can discuss.

Certifications such as ISO 27001 can be useful signals. They are not a substitute for control.

Security topics to discuss before sprint one

Ask the vendor to explain:

Identity and access management
Privileged access controls
Secrets handling in development and CI/CD
Dependency scanning and vulnerability management
Logging and evidence retention
Data residency and data location
Incident detection and response
Backup and recovery
Environment separation
Secure code review practices
How security issues enter the delivery backlog
How developers are trained on secure development

The goal is not to demand a perfect answer to every question.

The goal is to understand whether security exists inside daily delivery or only inside policy documents.

Red flags in security conversations

Be careful if:

Nobody can explain how secrets move through the pipeline
Incident response sounds like a shared support mailbox
Access control is handled informally
Production access is not clearly limited
Logs and audit evidence are not discussed
Security is described as something the client will “approve later”
The vendor relies only on generic compliance claims

For regulated fintech products, security must be designed into delivery from the beginning.

Keep code, pipelines, and knowledge visible

In fintech outsourcing, slow delivery is not always the biggest risk.

The bigger risk is losing control over code, environments, delivery knowledge, and operational resilience.

If the vendor owns the repo, pipeline, observability, runbooks, and release knowledge, you do not have a flexible delivery partnership. You have a dependency.

That might be acceptable for a narrow, low-risk service. It is dangerous for a regulated product or a core fintech capability.

A stronger operating model

A healthier setup usually includes:

Code stored in the client’s repository
Client-visible CI/CD pipelines
Client-controlled access policies
Shared observability
Clear architecture notes
Test evidence available to the client
Documented release procedures
Runbooks for operational scenarios
Regular knowledge-transfer sessions

This does not slow good vendors down.

It makes delivery safer. It also makes the relationship more balanced.

Design the exit before you need one

Many teams avoid exit planning because it feels pessimistic.

That is a mistake.

Exit planning is not a sign of distrust. It is a sign of mature outsourcing governance.

DORA requires documented exit plans for relevant ICT contracts. The PRA expects firms to consider timing, cost, resources, and risk triggers during the pre-outsourcing phase.

A fintech company should know how it would continue delivery if the vendor relationship changed.

Define exit readiness before sprint one

Before work begins, agree on:

Where code will live?
Where technical documentation will live?
What must be documented every sprint?
How knowledge transfer will happen?
Who owns operational runbooks?
Which events trigger transition planning?
How access will be removed at the end of the contract?
How unfinished work will be handed over?
What support the vendor provides during transition?

If a vendor is confident in its delivery quality, this should not be a controversial conversation.

Agree on delivery metrics that measure quality

If a vendor is measured only by hours and headcount, you should not be surprised when you get more hours and more headcount.

Those incentives are weak.

For fintech software outsourcing, delivery metrics should show how reliably the team moves valuable work into production.

Metrics to agree before sprint one

Consider tracking:

Lead time from idea to production
Deployment frequency
Change failure rate
Time to restore service
Defect escape rate
Cycle time by story type
Production incident trends
Test coverage for critical flows
Security issue resolution time
Backlog ageing for risk-related work

These metrics move the conversation away from activity and towards operating quality.

They also make vendor performance easier to discuss without emotion.

Use a pre-sprint scorecard

A simple scorecard can reveal whether you are dealing with a reliable fintech software development partner or a risky contractor.

Before sprint one, answer yes or no.

Pre-sprint outsourcing checklist

Have we met the named delivery lead?
Have we spoken with the people who will own security, QA, DevOps, and architecture?
Have we calibrated seniority through a real working session?
Do we know who will actually write the code?
Is the subcontracting chain visible?
Will the work live in a client-visible environment?
Can the vendor explain secure SDLC practices in practical terms?
Are data location, access controls, and evidence retention documented?
Do audit and access rights work in practice, not only on paper?
Is there a draft exit and knowledge-transfer plan?
Are delivery KPIs defined beyond hours and headcount?

How to read the scorecard

If you have zero to three yes answers, you likely have the wrong vendor or the wrong control model.

If you have four to six yes answers, proceed only if the scope is non-critical and supervision is strong.

If you have seven to eleven yes answers, you may have a reasonable starting point, provided the contract supports the operating model.

The contract should match how delivery will actually work. If the contract promises control but the delivery setup removes visibility, the risk remains.

The real outsourcing decision in fintech

This is not an argument against outsourcing.

Strong external engineering teams can help fintech companies move faster, access rare expertise, challenge weak assumptions, and bring structure to complex product delivery.

The real decision is not whether to use a contractor.

The real decision is whether the contractor’s delivery model helps you maintain control over risk, knowledge, software quality, and operational resilience.

For banks, financial institutions, fintech startups, product owners, and enterprise teams, that control matters as much as velocity.

Before sprint one, look beyond the sales deck. Meet the real team. Test how they think. Understand the supply chain. Make security operational. Keep environments visible. Design the exit. Measure quality, not effort.

A reliable fintech software outsourcing partner will welcome that level of clarity.

A risky one will avoid it.

Gemma 4 Is the First Open Model I’d Actually Recommend to a Client

19 Mayıs 2026 yazar Mert

This is a submission for the Gemma 4 Challenge: Write About Gemma 4

A client asked me three months ago whether they could build an AI feature into their SaaS product without depending on OpenAI.

Their concerns were reasonable. Vendor lock-in. Data leaving their servers. Cost unpredictability at scale. The risk of a pricing change breaking their unit economics overnight.

Six months ago, my honest answer would have been: “Not really. The open models are close, but not close enough for a production product recommendation.”

After spending a week with Gemma 4, my answer has changed. This is the first open model family where I’d say yes without caveats.

Here’s why, and what you need to know to actually use it.

The Questions That Actually Matter for Product Builders

Most Gemma 4 coverage asks: “Is it as good as GPT-4o?”

That’s the wrong question if you’re building products.

The questions that matter are:

Can I ship it commercially without legal risk?
Will it run on hardware my client already owns or can afford?
Is the quality actually good enough for the specific task I need?

For the first time with an open model family, Gemma 4 answers all three with a clear yes. Let me walk through each.

Question 1: Can I Ship It Commercially?

Every previous Gemma release came with Google’s custom Terms of Use. You could experiment, you could research, but commercial deployment required reading fine print and accepting restrictions that most small teams didn’t bother to fully understand.

Gemma 4 is Apache 2.0.

That means: use it commercially, fine-tune it, redistribute it, build products on it, charge money for those products. No MAU limits. No usage restrictions. No negotiating access agreements. The same license as React, TensorFlow, and half the software stack your product already runs on.

For a client project, this is the difference between “interesting prototype” and “production feature.” Apache 2.0 removes the legal conversation entirely.

Question 2: Will It Run on Available Hardware?

This is where Gemma 4’s architecture actually shines, and it requires understanding what Google shipped.

Gemma 4 is not one model. It’s four, each designed for a different hardware reality.

E2B — For Edge and Mobile

The “E” stands for effective parameters. Using Per-Layer Embeddings (PLE), the model runs at approximately 2B effective parameters during inference despite the weight file being larger. It runs fully offline on phones, Raspberry Pi, and Jetson Nano devices. Google’s benchmarks show it hitting roughly 7.6 tokens per second on a Raspberry Pi 5.

Hardware requirement at 4-bit: ~5 GB RAM
Best for: Mobile apps, edge deployments, offline-first features, anything battery or privacy sensitive

E4B — For Most Laptops

Same “E” naming, around 4.5B effective parameters at runtime. This is the entry point for most developers. Runs comfortably on a laptop with 8-12 GB RAM, supports audio and image input, and delivers noticeably better reasoning than E2B.

Hardware requirement at 4-bit: ~8 GB RAM
Best for: Developer laptops, local coding assistants, document processing, any standard dev machine

26B A4B — The Sweet Spot for Desktops and Servers

This one deserves a longer explanation because the naming trips people up constantly.

The model has 26 billion total parameters. But it’s a Mixture of Experts (MoE) architecture, specifically 8 of 128 experts plus 1 shared expert activate per token. The “A” in A4B means active parameters during inference: 3.8 billion.

What this means practically: you get quality close to a 26B dense model at roughly the inference speed of a 4B model.

On the LMArena leaderboard, it scores 1441 Elo. The 31B Dense scores 1452. An 11-point gap that is invisible in most real-world tasks.

One thing that trips people up: all 26 billion parameters still need to load into memory for fast expert routing. You cannot skip the memory cost just because inference is fast. Budget ~18 GB at 4-bit, not 4 GB.

Hardware requirement at 4-bit: ~18 GB RAM
Best for: Desktop workstations, local AI servers, coding agents, RAG pipelines, any machine with 16+ GB

31B Dense — Maximum Quality

Every token activates all 31 billion parameters. Slower, but the quality leader. Currently ranked #3 on the Arena AI open model text leaderboard, with the 26B A4B at #6, both outcompeting models 20x their size. Also the cleanest choice for fine-tuning because the dense architecture gives predictable gradient flow during training.

Hardware requirement at 4-bit: ~20 GB RAM (RTX 3090 or 4090 handles this)
Best for: Fine-tuning, maximum quality requirements, cloud GPU deployments

The Quick Decision Table

Your Hardware	Pick This
Phone / Raspberry Pi / embedded device	E2B
Laptop, 8-12 GB RAM	E4B
Desktop or Apple Silicon Mac, 16+ GB	26B A4B
Workstation / Cloud GPU, 24 GB+ VRAM	31B or 26B A4B
Fine-tuning a custom model	31B
Android app via AICore Developer Preview	E2B or E4B

Practical rule: start one size smaller than you think you need. A model that runs is more useful than one that swap-thrashes or crashes on load.

Question 3: Is the Quality Good Enough?

For the specific tasks most product builders need, yes.

Gemma 4 made two architectural improvements that matter for real usage, not just benchmarks.

Long context that actually works. Previous open models advertised long context but fell apart in practice when information was buried deep in a document. Google’s benchmark numbers show the 31B went from 13.5% to 66.4% on multi-needle retrieval tests. That is a functional change, not an incremental one. Passing a full codebase or a long contract in a single prompt now works in ways it didn’t before.

Native function calling. Tool definitions go directly into the model via apply_chat_template. No prompt engineering tricks to force JSON output. No wrapper libraries to parse responses. The model returns structured tool calls that you parse and execute. This is what makes local agentic workflows practical.

One thing to keep in mind: the training data cutoff is January 2025. For tasks requiring current information, pair it with RAG or a web search layer. This isn’t a weakness unique to Gemma 4, it’s true of any local model, but it’s worth planning for.

Running It: The Complete Setup

Via Ollama (Fastest Path)

# Install Ollama from ollama.com, then:
ollama pull gemma4:e4b     # ~8 GB
ollama pull gemma4:26b     # ~16 GB
ollama pull gemma4:31b     # ~20 GB

# Start chatting
ollama run gemma4:e4b

Critical step people skip: Ollama defaults to a 4,096 token context window regardless of what the model supports. It silently truncates from the beginning when the limit is hit, no error, no warning, no indication anything went wrong. If the model seems to “forget” earlier parts of a conversation, this is why.

Set it explicitly at runtime:

ollama run gemma4:e4b -p num_ctx=32768

Or in the API:

curl http://localhost:11434/api/chat 
  -d '{
    "model": "gemma4:e4b",
    "options": { "num_ctx": 32768 },
    "messages": [
      { "role": "user", "content": "Your prompt here" }
    ]
  }'

To make a permanent preset with your preferred context size:

FROM gemma4:e4b
PARAMETER num_ctx 32768

ollama create gemma4-e4b-32k -f Modelfile
ollama run gemma4-e4b-32k

Integrating with a Node.js / MERN Project

Ollama exposes an OpenAI-compatible API. If you have an existing project using the OpenAI SDK, the change is two lines:

import OpenAI from "openai";

// Before: const client = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });
// After:
const client = new OpenAI({
  baseURL: "http://localhost:11434/v1",
  apiKey: "ollama", // required by the SDK, ignored by Ollama
});

async function reviewCode(code) {
  const response = await client.chat.completions.create({
    model: "gemma4:26b",
    messages: [
      {
        role: "system",
        content: "You are a senior backend developer. Review code for bugs, security issues, and performance problems. Be specific and concise."
      },
      {
        role: "user",
        content: `Review this code:nn```jsn${code}n````
      }
    ],
    max_tokens: 800,
  });

  return response.choices[0].message.content;
}

This works because Gemma 4 has native system prompt support. In Gemma 3, the system role needed workarounds and inconsistent handling. Now it works exactly as the OpenAI API spec defines it.

Function Calling (The Agentic Part)

For tool use, the correct approach is AutoProcessor with apply_chat_template, passing tools as an argument. This is documented in Google’s official function calling guide at ai.google.dev.

from transformers import AutoProcessor, AutoModelForImageTextToText
import torch

MODEL_ID = "google/gemma-4-E4B-it"

model = AutoModelForImageTextToText.from_pretrained(
    MODEL_ID,
    device_map="auto",
    torch_dtype=torch.bfloat16
)
processor = AutoProcessor.from_pretrained(MODEL_ID)

# Define your tools as JSON schema
tools = [
    {
        "name": "search_docs",
        "description": "Search internal documentation for an answer",
        "parameters": {
            "type": "object",
            "properties": {
                "query": {
                    "type": "string",
                    "description": "The search query"
                },
                "max_results": {
                    "type": "integer",
                    "description": "Number of results to return"
                }
            },
            "required": ["query"]
        }
    }
]

messages = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "How do I reset my password in the admin panel?"}
        ]
    }
]

# Pass tools directly to apply_chat_template
inputs = processor.apply_chat_template(
    messages,
    tools=tools,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    add_generation_prompt=True,
).to(model.device)

input_len = inputs["input_ids"].shape[-1]
output = model.generate(**inputs, max_new_tokens=200)

response = processor.decode(output[0][input_len:], skip_special_tokens=True)
print(response)
# Returns a structured tool call you parse, execute, then pass the result back

For basic text inference with the pipeline API, Gemma 4 uses the "any-to-any" task type:

from transformers import pipeline

pipe = pipeline(
    task="any-to-any",
    model="google/gemma-4-E4B-it",
    device_map="auto",
    dtype="auto"
)

messages = [
    {
        "role": "user",
        "content": [{"type": "text", "text": "Summarize this in 3 bullet points: [your text here]"}]
    }
]

output = pipe(messages, max_new_tokens=300)
print(output[0]["generated_text"][-1]["content"])

Note: the task type is "any-to-any", not "text-generation". Using the wrong task type is a common mistake when adapting older Gemma 3 code.

What I’d Actually Use Each Model For in Client Projects

Based on a week of real testing, here’s how I’d map these to actual product use cases:

E2B / E4B: Offline-first mobile features. Smart reply suggestions. On-device document summarization where the user’s data cannot leave the device. HIPAA-adjacent use cases where cloud AI is simply not an option.

26B A4B: This is the model I’d use for most backend AI features. A support ticket classifier. An internal knowledge base Q&A system. A code review step in a CI pipeline. A document analysis feature for a B2B SaaS tool. It’s fast enough for interactive use, good enough for most tasks, and runs on hardware a small company can own outright.

31B: Fine-tuning for a specific domain. A legal document analyzer trained on a client’s specific contract types. A customer support model trained on a company’s historical tickets. Cases where you need the base model quality before fine-tuning and where inference latency is less critical than output quality.

The Honest “When Not To Use It” Section

Gemma 4 doesn’t replace every cloud AI use case, and claiming otherwise would be overselling it.

If your team has no GPU infrastructure and no plans to build any, Google AI Studio gives you hosted access to the 31B and 26B A4B with zero setup. Start there.

If you need current information beyond January 2025, plan for RAG from day one. The model itself won’t have it.

If you’re building something that needs the absolute frontier of reasoning capability, GPT-4o and Claude still lead on the hardest tasks. Gemma 4 is exceptionally strong for its size, but the top closed models still have an edge on the most complex multi-step reasoning.

Where Gemma 4 wins is the space between “toy project” and “frontier reasoning.” That space covers the majority of real product features most development teams are actually building.

Where the Open Model Story Is Heading

Gemma 4’s real significance isn’t any single benchmark number. It’s that the capability-to-hardware ratio has crossed a threshold.

A model that runs on a 16 GB desktop, handles 256K context that actually retrieves correctly, does structured tool calling natively, processes images and audio, supports 140+ languages, and ships under Apache 2.0 — that’s a different product decision than anything that existed a year ago.

The 400 million Gemma downloads across all versions, and the 100,000+ community-built variants that followed, suggest developers were waiting for exactly this. A model good enough to build on, free enough to ship, small enough to run.

Gemma 4 is that model. For a lot of product builders, this is the point where “local AI” stops being a research interest and starts being a line item in a project estimate.

Start Here

No setup: Google AI Studio — hosted 31B and 26B A4B, free to start
Local CLI: Ollama — ollama run gemma4:e4b and you’re running
Local GUI: LM Studio — point and click, no terminal needed
Direct weights: Hugging Face — google/gemma-4-31B-it, google/gemma-4-26B-A4B-it, google/gemma-4-E4B-it, google/gemma-4-E2B-it

If you’re starting today: pull gemma4:e4b on Ollama, set num_ctx to at least 8192, and test it on a task you actually need solved. That’s more useful than reading another benchmark comparison.

What are you building with Gemma 4? Drop your use case in the comments. Especially interested in anyone using E2B or E4B for on-device features.

Home Robot Safety Is All About Relationships

19 Mayıs 2026 yazar Mert

The International Standards Organization (ISO) is updating its 12-year-old safety requirements for personal care robots. A lot has happened since the last revision, both on the technology side and with researcher’s understanding of safety for humans collaborating with domestic robots. The proposed ISO update addresses hazard identification, risk assessment, and different use scenarios. It does not, however, set limits, propose testing methods, or have enforcement mechanisms that might address the complexities of human-robot collaboration. And that is a problem, argues technology policy researcher Jae-Seong Lee of the Electronics and Telecommunications Research Institute in Daejoeon, South Korea.

IEEE Spectrum: Why is the next revision of ISO 13482 a big deal?
Jae-Seong Lee: The standard is moving into final approval at a moment when domestic humanoid robot makers are shifting from lab prototypes to products aimed at real homes, real caregivers, and real families. That matters because the standard does more than specify geometry and impact limits: it helps define what counts as acceptable robot behavior in the messy world of everyday life.

What is the core engineering problem?
Lee: It is not simply whether a robot can avoid collisions or detect people in its path. The harder problem is that human-robot interaction is bidirectional. The robot changes what the human does, and the human changes what the robot perceives and does next. In other words, safety is not a fixed property of the machine alone; it emerges from the relationship.

Isn’t that already covered by current safety standards?
Lee: Only partially. ISO 13482 addresses personal care robots through hazard identification, risk assessment, and intended use scenarios, and related guidance acknowledges non-contact hazards such as unpredictability and incorrect autonomous decisions. But it stops short of binding compliance criteria, test methods, or enforcement mechanisms for the hazards produced by the human-robot relationship.

The technical community understands bidirectional coupling, and the standards framework acknowledges relevant hazards, but no current standard fully converts that knowledge into enforceable rules for domestic autonomy.—Jae-Seong Lee

Why can’t engineers just better define a robot’s operating envelope?
Lee: Because the value proposition of a domestic humanoid depends on operating in uncontrolled environments. A robot that is safe only in standardized rooms, with healthy adults, under well-defined conditions is not really a domestic humanoid at all.

In industrial robotics, designers can usually bound the task, the workspace, and the population. In a home, the robot must adapt to elderly residents, children, visitors, pets, clutter, tight spaces, and fluctuating human behavior. Those aren’t edge cases. Those are the baseline. Tightening the domain to be more like that of factory robots would make the home robots less useful.

The proposal mentions training data. Why does that matter?
Lee: Because the data already reflect the diversity of domestic life. Companies building humanoid training datasets are reportedly sending paying contract workers around the world to record their chores in ordinary settings. That means the robots will be trained on real-world variability, not sanitized demonstrations. The safety problem is therefore in the composition of the entire human-robot system, not in any one component.

What is the standards gap?
Lee: The gap is governance. The technical community understands bidirectional coupling, and the standards framework acknowledges relevant hazards, but no current standard fully converts that knowledge into enforceable rules for domestic autonomy. What is missing is a way to specify safe behavior across the full range of human conditions the robot will actually encounter.

What’s also missing is a decision about who gets to decide whose behavior counts as normal. Whose gait sets the baseline? Whose is an acceptable risk threshold? Whose definition of safe judgment gets written into the requirement language? Those are value judgments, not purely engineering ones. A standards committee cannot avoid choosing a normative reference point; it can only decide whether that choice is explicit and inclusive.

Who could help answer those questions?
Lee: The proposal argues that the people most affected by domestic humanoids are not systematically represented in the working groups shaping the standard. It points especially to older adults, who are often the primary intended users of domestic care robots, yet whose movement patterns and cognitive states are not directly embedded in the standards process.

In other words, this revision acknowledges the hardest problems but pushes unresolved issues into advisory language, non-binding guidance, or future revision scopes. That can be useful, but it also delays the real question: what counts as safe relational behavior in the home?

What are the stakes?
Lee: The risk is not only injury, though that is the obvious concern. The deeper risk is that safety assumptions get baked into products and standards before the market, regulators, and users have a chance to question them. Once deployment patterns harden, it becomes much harder to revise the baseline.

What should the engineers on the standards bodies do about it?
Lee: The engineers on the standards body should ask not just, “What are the robot’s outputs, and do they stay within safe thresholds?” but “What states does this robot engage with, and does that engagement remain safe across the full range of those states?” That shift sounds subtle, but it changes the design brief. It moves safety from machine-centric measurement toward system-level relational assurance.

Domestic humanoid safety cannot be solved by machine engineering alone. It requires a framework that treats the human not as background noise, but as part of the system, part of the definition of the safety envelope.