AI

Stak Energy Plans $500M AI Data Center on Alaska’s North Slope

Published

15 hours ago

May 27, 2026

Stak Energy, an Alaska-based startup, has filed for state approval to build a $500 million AI data center on the North Slope, covering more than a square mile of tundra off the Dalton Highway and targeting up to 3 gigawatts of power drawn entirely from natural gas that currently sits stranded beneath Arctic oil fields. On May 12, the Alaska Department of Natural Resources issued a preliminary finding that the project is in the state’s “best interest,” opening a public comment period through June 15 before any land lease can proceed.

The pitch rests on geography: every friction point strangling AI data center buildouts in the Lower 48, from spiking electricity prices to water shortages and community opposition, largely disappears at 70 degrees north latitude. Whether Stak can translate that geographic logic into a funded, turbine-equipped, customer-signed facility by 2028, its stated target for first operations, is the question it has yet to answer publicly.

Project Aaka and the Square-Mile Scope

Stak calls the development Project Aaka, named for the Iñupiaq word for “grandmother.” At full build-out it would place multiple buildings across approximately one square mile of land some 25 miles south of the North Slope’s major Prudhoe Bay infrastructure, connected to petroleum fields by a newly constructed natural gas pipeline. State documents show the power plant feeding those buildings could consume more than twice the natural gas that all of urban Alaska burns annually for electricity generation and combined home and commercial heating.

At a maximum of 3 gigawatts, the project would sit in the same general tier as the largest AI campuses currently under development in the Lower 48. Meta’s Hyperion campus in Louisiana is targeting 7.46 gigawatts; Microsoft and Chevron are building a five-gigawatt gas plant in West Texas; Google is partnering with Crusoe Energy on a 933-megawatt facility in North Texas. Stak’s ceiling is more modest, but its fully off-grid, self-powering design is structurally different from all of them.

Feature	Stak Energy (North Slope)	Typical Lower 48 Hyperscale
Power capacity	Up to 3 GW	0.5 to 7+ GW
Power source	Stranded natural gas, fully off-grid	Grid-tied, mixed sources
Cooling method	Arctic ambient air	Water cooling towers
Estimated water use	90% below industry norm	Standard industrial volume
Average site temperature	12°F annually	40°F to 65°F
Local ratepayer grid impact	None	Raises local electricity rates

Stak officials declined to respond to specific questions about the proposal. The company’s prepared statement called the lease filing “an important milestone for anchoring Alaska as America’s at-scale energy solution” and described the company as committed to “responsible development” of Alaska’s economy. No anchor tenants or committed investors are named in any public document filed with the state.

Stak Energy plans $500 million AI data center on Alaska North Slope.

Three Reasons Alaska Beats Virginia

Northern Virginia holds most of the world’s existing data center capacity, but the region’s grid is straining under the load, its water tables are under pressure, and community opposition to new campuses grows louder at planning meetings across the state. Stak’s geographic argument comes down to three physical advantages that no warmer-latitude market can replicate:

Stranded natural gas with no competing demand. North Slope oil fields hold large reserves of associated gas that petroleum companies have historically left in place because no pipeline connects them to buyers. Stak would face virtually no price competition for that supply, at least at project launch.
Arctic air cooling. With an average annual temperature of 12°F, the campus would use outside air to cool its servers rather than relying on water towers. The lease application says that approach cuts water consumption by 90% or more compared with industry norms, removing a key source of local opposition that has plagued data center projects elsewhere.
Complete grid independence. The facility would not connect to Alaska’s main power grid, creating no new load for Anchorage or Fairbanks ratepayers. Data centers in Virginia, Texas, and Oregon have faced backlash precisely for driving up household electricity bills.

Antony Scott, a former commercial petroleum analyst for the Alaska Department of Natural Resources, told Northern Journal that the off-grid structure clears a significant political obstacle. “The issue of data centers and the effect on normal humanity’s electricity bills is causing real angst,” Scott said. On the North Slope, he added, “we avoid all of that. You can just step into this friendly environment.”

The Company and Its Politically Connected Hires

Sparrow Mahoney, Stak Energy’s founder and chief executive, grew up in Alaska and attended Wasilla High School. She launched the company roughly a year ago with a far narrower goal: convert stranded North Slope gas into electricity and use it to mine Bitcoin. Prior involvement in the Iditarod’s push into cryptocurrency was part of her startup background, according to Alaska Public Media. The pivot to large-scale AI and cloud computing has transformed both the company’s pitch and the scale of capital it now needs to raise.

Alaska is the only place this makes sense, long term, for the industry.

Mahoney said this at the May 2026 Alaska Sustainable Energy Conference, where she cited U.S. Department of Energy figures showing the country needs 100 gigawatts of new data center capacity by 2030. To staff up for that ambition, Stak made a series of politically connected hires. John Boyle, who served as natural resources commissioner under Governor Mike Dunleavy, joined the company, as did Jim Shine, a former special assistant at the same department.

Both bring familiarity with the land-use and permitting process the project must work through, including a federal Clean Water Act authorization to build a gravel pad that documents indicate would be nearly twice the size of the one at ConocoPhillips Alaska’s Willow oil field. Fundraising is running through McKinley Alaska Private Investment, an Anchorage-based firm. How far along Stak is toward its $500 million target is not public. Stak Energy’s project materials describe the venture as building “the secure, scalable energy foundation for American dominance,” a framing that fits the current federal energy policy environment comfortably.

The AI Power Crunch That Sent Stak North

Stak’s proposal didn’t arrive in a vacuum. Across the United States, AI training and inference workloads are consuming electricity fast enough to strain every major data center market simultaneously. Former U.S. Senator Kyrsten Sinema of Arizona, now co-chair of the AI Infrastructure Coalition, put it plainly at a May 2026 energy conference: “The number one bottleneck that we face is energy.”

6.7% to 12%: the share of all U.S. electricity that data centers could consume by 2028, per a U.S. Department of Energy projection cited in Mitsubishi Power’s sector outlook
40%: share of AI data centers projected to face electricity deficits by 2027, per Gartner
5 to 7 years: current lead times for new combined-cycle gas turbines, per S&P Global

That turbine shortage affects every natural-gas data center build in the country, not just Stak. GE Vernova, Siemens Energy, and Mitsubishi Heavy Industries collectively supply most of the world’s new gas turbines, and all three are reporting delivery schedules stretching well past 2029. Some manufacturers have reportedly stopped accepting new orders without committing to delivery timelines. U.S. data centers are expected to add between 3 and 6 billion cubic feet per day of new natural-gas demand by 2030, meaning competition for both fuel and turbines will intensify before it eases.

The Gaps the Lease Papers Don’t Fill

No Gas Supply Deal on File

The lease application the Alaska DNR reviewed was thorough on engineering concepts, but contained a material gap: at the time Stak filed it, the company had not struck a firm supply agreement with any petroleum operator on the North Slope. The proposed pipeline connecting the campus to a gas source could run anywhere from 25 to 90 miles, a range that implies Stak has not settled on a specific field or a specific seller.

Antony Scott told Northern Journal that the range gives the situation away. “That means they don’t have a gas supply,” he said. North Slope operators are presumed willing to sell associated gas they currently cannot monetize, but presumed willingness is not a signed contract with price and volume terms. A binding supply agreement is the project’s most pressing next milestone.

Turbine Timing vs. a 2028 Start

Stak says it wants initial operations running by 2028. That target runs directly into the gas turbine shortage. S&P Global’s analysis of the U.S. gas turbine supply crunch puts current lead times at as long as seven years, and some manufacturers have effectively closed order books until 2028 without committing to delivery dates. Getting into a manufacturer’s queue today would not guarantee turbines arrive before 2030 under current backlog conditions.

Stak has not disclosed whether it has placed orders with any manufacturer or secured production slots. For a facility with a 2028 first-power target, turbine procurement is the item that every other schedule depends on. Without it, project timelines are aspirational rather than contractual.

Investors and Customers

The company has said it is raising capital through McKinley Alaska Private Investment, but has not named institutional backers, disclosed how much equity is committed, or confirmed any AI operator as a tenant. Northern Journal reporter Nathaniel Herz found that Stak had not been forthcoming about confirmed customers or investors. The lease application, technically detailed and described by energy experts as thorough, stops short of naming any offtake agreements or committed equity investors.

That profile, a credible site with professionally assembled permitting documents and politically connected advisors but no confirmed anchor tenant, is common at the early stage of large infrastructure projects. It is also the stage at which the most ambitious ones most often stall indefinitely.

Carbon, Permitting, and the Path From Paper to Power

Running 3 gigawatts off natural gas creates a carbon liability that Stak’s potential tenants cannot easily ignore. Technology companies including Microsoft, Google, and Meta maintain formal emissions-reduction targets, and a fossil-fuel-only power source in a remote Arctic location without access to wind, solar, or existing carbon-capture infrastructure creates a sourcing problem for any tenant that must report against a net-zero commitment.

Stak says in its application that it is monitoring carbon capture and storage technology, but acknowledges that the North Slope’s geology for sequestration is poorly understood and that no capture infrastructure exists there yet. For the foreseeable future, each megawatt-hour the facility produces will carry a full fossil fuel emissions footprint.

Governor Dunleavy’s administration formally found the project in the state’s best interest on May 12, and the Alaska DNR public comment portal is accepting input on that preliminary decision through June 15 before any lease can be finalized. If Stak can show a signed gas supply agreement, confirmed turbine orders, and named tenants before the 50-year lease goes final, the project becomes a genuine competitor in the AI infrastructure race. If those three items are still open when a new Alaska governor takes office in 2027, the square mile off the Dalton Highway will be the most detailed unfunded concept in the state’s energy history.

AI

UC San Diego AI Foundation Model Predicts Cancer Treatment Response

Published

16 hours ago

May 27, 2026

Logan Pierce

Only about 8% of cancer patients whose tumors are sequenced end up matched to an FDA-approved therapy on the basis of genetics. That number reflects not a failure of sequencing technology – a standard clinical panel already finds roughly 11 distinct genetic alterations in the average tumor – but a fundamental shortage of tools for interpreting what most of those mutations mean for treatment choice.

Researchers at the University of California San Diego published a paper Monday in Cancer Discovery describing an AI foundation model called MutationProjector that reads a tumor’s full mutation landscape and predicts treatment response with accuracy matching or exceeding existing methods across several independent patient cohorts. The second-order implication gets less attention than the performance numbers: if a system like this holds up at scale, it changes the purpose of a sequencing panel from a narrow flag-list into a comprehensive treatment-strategy input.

The 92% Gap in Genomic Treatment Matching

8%: share of cancer cases successfully matched to an FDA-approved therapy by genetics alone, per the UC San Diego research team.
~11: average number of distinct genetic alterations identified in one tumor by a standard clinical sequencing panel.
30,000+: tumor genomic profiles across 10 solid cancer types used to train the foundation model.
1 gene: the basis for the vast majority of those successful genetic matches – standard treatment stratification typically hinges on a single biomarker at a time.

Genetic testing after a cancer diagnosis has become routine for good reason. The published study in Cancer Discovery notes that DNA sequencing panels flagging alterations in cancer-associated genes are relatively fast and low-cost, with a strong track record where validated biomarkers exist. The problem is that validated biomarkers currently cover a narrow slice of what a panel actually finds.

“Genetic sequencing is already routine in cancer care, but we still struggle to fully interpret the many mutations found in a patient’s tumor,” said Trey Ideker, PhD, professor of medicine at UC San Diego School of Medicine and director of the Big Data Institute at the University of Oxford, who led the research. Clinical treatment decisions based on genetics today still hinge mostly on a single gene, whether BRCA1, EGFR, or KRAS, while the other 10 alterations in the same report sit unused.

Two problems make those unused mutations hard to act on. First, many individual mutations are rare – appearing in only a tiny fraction of patients – which makes them nearly impossible to study through conventional clinical trials. Second, one mutation’s effect on drug response often depends on what other mutations are present in the same tumor. Standard biomarker testing checks each gene in isolation and misses interaction effects entirely.

AI model predicts cancer treatment response from tumor mutation profiles.

How the Model Encodes a Tumor’s Full Mutation Landscape

From Mutation Lists to Biological State

A standard biomarker approach scans for a known alteration, confirms its presence, and maps it to a drug. The UC San Diego team’s system works differently. It takes the full set of genetic alterations identified in a tumor and compresses them into what the researchers describe as a “compact representation” of the tumor’s biological state – a summary of which molecular pathways are likely disrupted, rather than a catalogue of individual mutations.

That representation is what the system uses to make treatment predictions. Instead of asking “does this tumor carry KRAS G12C?” as a binary flag, it asks what the combined mutation profile implies about how the tumor’s cellular machinery is functioning. The shift from mutation list to pathway state is the core architectural choice behind the model’s design, and it is what allows the system to produce an explanation alongside each prediction rather than a raw probability score.

The model belongs to a class called foundation models (AI systems pretrained on large, broad datasets and then adapted for specific downstream tasks with relatively small labeled samples). Foundation models underpin large language models in general AI applications; the UC San Diego version runs on tumor genomes rather than text, but the conceptual logic is the same: expose the model to enough examples across enough variation that it can detect patterns no single institution could find in isolation. A 2026 Nature Reviews Cancer perspective on machine learning and genomics notes that as large clinicogenomic datasets mature, such tools have a real opportunity to extract more information from sequencing data and generate therapeutic hypotheses for patients who currently have none.

Why Rare Mutations Become Detectable at Scale

Training on more than 30,000 tumor profiles across 10 solid cancer types gives the model exposure to mutations that no single institution sees often enough to study in a conventional clinical setting. “By pretraining on a large collection of tumors and integrating molecular network knowledge, MutationProjector can detect patterns that would be easy to miss with conventional biomarker approaches. That gives us a way to move from long lists of mutations toward a more functional understanding of the tumor,” said JungHo Kong, PhD, the study’s first author and a postdoctoral researcher in the Department of Medicine at UC San Diego School of Medicine.

The model integrates biological network data – information about how genes and proteins interact – alongside raw mutation profiles. That integration connects mutations to known biological pathways, giving the predictions interpretable biological grounding rather than a pure statistical correlation drawn from the training data.

Attribute	Conventional Single-Biomarker Approach	Foundation Model Approach
Input data	One or a few targeted gene flags	Full mutation profile from standard clinical panel
Treatment signal	Presence or absence of a specific mutation	Combined pathway-level biological state
Handles rare mutations	No – needs sufficient trial data per mutation	Yes – patterns learned across large population
Interpretability	Direct binary match or no-match	Pathway-level reasoning per prediction
Training scale	Small labeled datasets per biomarker	30,000+ tumor profiles, 10 cancer types
Biomarker discovery	Limited to pre-specified known targets	Can surface unexpected new associations

Bladder Cancer, Lung Cancer, Melanoma: Validation Across Independent Cohorts

The research team validated the AI tool across several independent patient cohorts – populations whose data were held out from training and tested separately, the standard check for whether a model’s patterns hold beyond the data it saw during training. Across cancers including bladder cancer, lung cancer, and melanoma, the tool matched or exceeded existing methods for predicting response to both immunotherapy and chemotherapy.

Peer-reviewed publication in Cancer Discovery, a journal of the American Association for Cancer Research, placed the results through external evaluation before they were made public. The key performance findings from the published paper:

The tool matched or exceeded state-of-the-art prediction accuracy in every independent cohort evaluated.
Immunotherapy response predictions held across multiple cancer types, not just one tumor category.
Chemotherapy response predictions also held, suggesting the approach generalizes across different treatment mechanisms rather than optimizing for a single drug class.
Multi-cohort validation separated the results from single-dataset overfitting, a common failure mode in earlier machine-learning oncology research.

Earlier attempts to apply machine learning to cancer treatment prediction have sometimes produced strong numbers on a training dataset and poor performance when tested elsewhere. Building the validation around independent cohorts was deliberate, and the consistency of results across cancer types and treatment classes is among the more credible aspects of the published work.

Two Biomarker Findings the Training Data Did Not Obviously Predict

Beyond the core accuracy results, the model identified associations that were not part of the researchers’ initial hypotheses – and at least one sits in tension with how current clinical sequencing guidance is constructed.

Mutations in the gene KMT2D were flagged as a predictor of sensitivity to immunotherapy. KMT2D encodes a histone methyltransferase involved in gene regulation. Prior research had associated KMT2 family mutations with immune checkpoint inhibitor response in pan-cancer analyses, but the finding had not been incorporated into standard clinical sequencing guidance. The model’s identification of KMT2D sensitivity, surfaced from patterns across the training population, points toward a candidate biomarker for prospective clinical study.

Our results suggest that tumor genome foundation models may help extend the clinical value of sequencing beyond a handful of well-known genes. This could support a more comprehensive and biologically grounded approach to precision oncology.

Trey Ideker made that comment in the UC San Diego release accompanying the paper’s publication on May 26.

The second finding is more operationally immediate. The joint presence of mutations in both SMARCA4 and STK11 emerged as a predictor of immunotherapy resistance. Prior published research had already linked SMARCA4 mutations to poor immunotherapy outcomes in lung cancer and STK11 mutations to resistance in non-small-cell lung cancer. What the model added was evidence that when both appear together in the same tumor, the resistance signal is stronger and more consistent than either gene alone. That co-occurrence pattern is precisely what single-biomarker testing misses: SMARCA4 mutant tumors have been documented as having poor immunotherapy response even at high tumor mutational burden, and the co-mutation picture with STK11 deepens that signal considerably.

Interpretability as the Clinical Bridge

A prediction model that cannot explain its reasoning has limited use in the clinic. Oncologists need to understand which biological features drove a prediction in order to weigh it against other patient factors and make a defensible treatment decision. That interpretability requirement was built into the foundation model’s design from the start, rather than added as a post-hoc visualization layer.

Because the tool maps mutation combinations to biological pathways rather than learning purely statistical correlations, the pathway disruptions it identifies can be read back as a form of explanation. A prediction of chemotherapy resistance, for instance, comes attached to information about which molecular pathways the model weighted most heavily – information a clinician or researcher can evaluate against existing biological knowledge, rather than being asked to trust an opaque score.

That design also shapes how the system can contribute to biomarker refinement over time. When a novel association surfaces, as with KMT2D, the pathway-level framing gives researchers a biologically grounded starting point for hypothesis generation, shortening the distance between a computational observation and a testable clinical question. A 2026 BJC Reports analysis on AI and precision oncology flags dataset diversity as a parallel challenge: training populations that underrepresent certain ethnic and geographic groups risk missing population-specific variations in tumor mutational burden and drug metabolism that affect therapy response – a limitation the UC San Diego team will need to address as the model expands.

Pancreatic Cancer, Prostate Cancer, Liquid Biopsies: The Expansion Roadmap

The published study covers 10 solid cancer types. Among the expansion targets explicitly named by the research team are pancreatic cancer, prostate cancer, and sarcomas – all with large unmet needs in treatment stratification and all representing tumor profiles the current training set did not include.

Beyond adding cancer types, the team wants to incorporate other data modalities. Medical imaging, transcriptomics (gene expression analysis, distinct from mutation detection), and electronic health records all carry treatment-relevant information that standard sequencing panels do not capture. Liquid biopsies – tests that detect circulating tumor DNA (ctDNA, fragments of tumor genetic material shed into the bloodstream) from a blood draw rather than a tissue sample – were specifically called out as a future application, particularly for early cancer detection before a solid tumor is large enough to biopsy directly.

The study listed co-authors from Lunit Incorporated, a Seoul-based AI company focused on oncology applications, alongside the UC San Diego team. That collaboration points toward the kind of multi-institutional data-sharing that expansion will require: access to large international genomic datasets of sufficient quality to train on rarer cancer types and data modalities that are harder to standardize across different health systems.

If that data access materializes and the model holds its performance in new tumor contexts, the current 8% treatment-match rate faces structural pressure from a direction the field has not previously had the tools to pursue. If the gains prove specific to the cancer types and treatment classes in the current training set, the tool joins a growing list of promising oncology AI research that still awaits the harder test of broad clinical deployment – where dataset diversity gaps and regulatory standards will matter as much as benchmark accuracy numbers do.

AI

ByteDance’s 300 Job Postings Reveal an AI Agent OS Battle

Published

16 hours ago

May 27, 2026

Logan Pierce

More than 300 job openings at ByteDance this spring set off a familiar tech-industry rumor cycle: the TikTok parent was building phones again. ByteDance denied it Tuesday, telling Chinese outlet PencilNews that reports of a restarted self-developed handset project were false, and the denial is almost certainly accurate. The postings span Android Framework engineering, Telephony RIL (Radio Interface Layer, the software bridge between Android and cellular hardware), chip-driver development, radio-frequency design, and edge-side AI deployment, all positions belonging to a company trying to own the operating-system layer that determines what AI agents can see, reach, and pay for on any smartphone. That is the strategic territory the entire global AI industry is now converging on.

What 300 Job Postings Signal About the AI Agent Race

The positions break into two clusters. About 83 are tagged to Doubao Mobile Assistant, ByteDance’s AI agent product. The remaining 236 sit under mobile OS development. Together, they cover Android Framework, HAL (Hardware Abstraction Layer, the interface between Android software and device-level hardware), Telephony RIL, power consumption and thermal management, foldable-screen adaptation, and long-term memory architectures for agents. Several postings carry production-readiness language that does not belong in research roles: IP68 waterproofing specifications, NPI (New Product Introduction, the factory-readiness process for consumer devices), and ODM/OEM mass-production collaboration terms.

None of that is chatbot vocabulary. These positions describe a company trying to own what AI agents can see, touch, and command before any third-party app gets to decide what is permitted. IDC, the technology market research firm, projects 147 million AI phone shipments in China for 2026, a 31.6% year-on-year increase, with 53 percent of the total smartphone market expected to carry next-generation AI capabilities by year-end.

The commercial context sharpens the ambition. QuestMobile, a Chinese mobile analytics firm, tracks Doubao as China’s dominant AI application, ahead of rivals including DeepSeek. Yet despite that position, Doubao still depends on external platforms for every real-world transaction. Book a flight and the agent needs a travel platform. Order food and it requires Meituan’s permission. Send money and it needs Alipay or WeChat Pay. Each of those platforms can, and does, close the door.

345 million monthly active Doubao users as of March 2026, per QuestMobile
300+ ByteDance job openings for mobile OS and edge-AI roles as of late May 2026
147 million AI phone units IDC projects will ship in China in 2026, a 31.6% year-on-year increase
53% of China’s smartphone market IDC forecasts will be next-generation AI phones by year-end

ByteDance Doubao AI agent mobile operating system strategy and the system control battle.

The Nubia M153 Test and the Ecosystem Blockade

Capabilities the Prototype Demonstrated

On December 1, 2025, ByteDance and ZTE Corp unveiled the Nubia M153, a 6.78-inch Android prototype priced at 3,499 yuan (approximately $495) and limited to roughly 30,000 units. The device sold out on its first day, with secondary-market prices climbing more than 40 percent above retail. Positioned as an engineering prototype rather than a consumer flagship, it was designed to test one question: could an AI agent embedded at the OS level navigate Chinese apps and complete multi-step tasks without the user tapping a single screen?

Hands-on testing showed it could. The Doubao assistant, activated by a physical side button or voice command, identified a product from a social media post, compared prices across multiple e-commerce platforms, and navigated app interfaces using visual recognition rather than pre-built API connections. Ask it to play a specific podcast episode and it would open the app, search, find the episode, and begin playback without human input. Li Liang, vice president of Douyin Group, described the ZTE partnership on December 9, 2025 as “the starting point for the development of AI phones,” adding that “regardless of whether this attempt succeeds or not, AI is undoubtedly the future.”

The Walls That Super Apps Built

Within days of the launch, major Chinese platforms began pushing back. WeChat login showed abnormalities on the device. Taobao and Alipay raised access restrictions. Meituan’s food delivery flow rejected automated inputs because fraud-detection systems were not built for OS-level agents clicking through their interfaces. ByteDance preemptively suspended banking, payment, and gaming automation on December 5, 2025, describing the move as a “necessary step to ensure the technology has a more solid and far-reaching future.” By December 10, some Alibaba apps partially lifted restrictions after ByteDance disabled corresponding Doubao capabilities for those platforms.

The reasons were not purely technical. An executive source at a major e-commerce company told Japanese financial outlet Nikkei Asia that the real concern was control: an AI agent routing users to the cheapest platform cuts through the closed economic loop each super app depends on for traffic, recommendations, and transaction data. A financial app executive raised the accountability question directly: if a user instructs Doubao to transfer money and something goes wrong, who bears responsibility?

Doubao Capability	Status During December 2025 Test
Cross-platform price comparison	Functional
Podcast playback via voice and UI navigation	Functional
Ticket booking via Douyin integrations	Partially functional
Taobao automated ordering	Blocked by platform
Meituan food delivery flow	Blocked by fraud detection
Banking and payment operations	Suspended by ByteDance, Dec 5, 2025
WeChat login	Abnormalities reported during testing
Gaming rank automation	Suspended by ByteDance, Dec 5, 2025

ByteDance’s Unresolved Hardware Problem

ByteDance has attempted consumer hardware across nearly every major category over the past six years, and the results trace the same structural weakness: world-class software, essentially no offline retail presence in China. The Nubia M153’s app blockade is the newest manifestation of that gap, not its origin.

Pico Interactive, the VR headset company acquired in 2021 for approximately 9 billion yuan, received an additional 10 billion yuan by 2022, bringing total investment to over 20 billion yuan. Consumer sales fell well short of targets and the business underwent its largest-ever restructuring in late 2023, with the consumer division significantly scaled back.
ByteDance acquired part of the Smartisan Technology team and patents in 2019, following the struggling phone startup’s collapse. The Nut Pro 3 and Nut R2 devices that followed sold fewer than 100,000 units combined on JD.com and Taobao before the Nut phone line was closed in 2021.
A smart lamp product was disrupted by China’s regulatory environment for education hardware. Several audio hardware projects were discontinued before reaching consumers.
Doubao earphones and AI glasses remain in development, with the glasses team reportedly drawing on engineers recruited from Meta’s hardware division. Neither product has yet reached commercial scale.

The channel deficit explains why hardware remains structurally difficult for ByteDance. Huawei operates more than 10,000 offline retail stores across China. Xiaomi Home exceeds 16,000 locations. OPPO and vivo together maintain hundreds of thousands of distribution terminals. ByteDance has no equivalent physical retail infrastructure, which means any device it sells must win on spec and online visibility alone, without the floor staff who explain, demonstrate, and convert browsers into buyers.

The Entrance Everyone Wants and Nobody Controls

ByteDance is not the only company that has concluded the mobile OS layer is the prize. The question of who owns the AI agent stratum on a smartphone has become the central competitive question across the global technology industry, with several players that have far stronger hardware positions working on their own answers simultaneously.

AI agents will be foundational to the evolution of super apps, with success depending on deep integration across payments, logistics, and social engagement.

Charlie Dai, vice president and principal analyst at Forrester Research, made that observation to CNBC in January 2026, naming the three capabilities an agent needs to be genuinely useful at scale. Those three are the same chokepoints Doubao cannot currently cross without cooperation from WeChat, Meituan, and Alipay.

The bind is structural and competitive at once. According to Nikkei Asia’s December 2025 reporting, the reason ByteDance chose ZTE’s Nubia as its hardware partner was that the major manufacturers, including Huawei, Xiaomi, OPPO, and Honor, are all building proprietary AI assistants and would not grant a rival’s AI the system-level permissions that make an agent useful. ZTE had the commercial motivation to accept deeper collaboration precisely because it does not hold a dominant market position. The second-generation Doubao phone is again being developed with ZTE’s Nubia, with ByteDance simultaneously negotiating with OPPO, vivo, and smaller manufacturers to embed Doubao as a licensed system assistant on devices they control.

OpenAI, Apple, and the Pursuit of a Clean-Slate Device

Outside China, the same battle is developing on a longer timeline. Supply-chain analyst Ming-Chi Kuo of TF International Securities reported in late April 2026 that OpenAI was accelerating its first AI agent phone, with mass production now targeted for the first half of 2027, pulled forward from a prior 2028 estimate. Per Kuo’s April 26 supply-chain analysis, the device would be built around a customized MediaTek chip with Luxshare Precision Industry, a Chinese electronics manufacturer, as manufacturing partner, with combined shipments projected at around 30 million units across 2027 and 2028.

OpenAI’s theoretical advantage is that it has no legacy app ecosystem to protect. A device designed around agents rather than app icons would not need to negotiate screen permissions with Taobao or Meituan because those platforms would not be part of the interaction model. The disadvantage is the same one ByteDance already encountered: the major apps users depend on were not built for an external AI to operate them, and their owners have commercial reasons to keep it that way.

Company	AI Terminal Strategy	Hardware Path	Stage as of Mid-2026
ByteDance (Doubao)	OS-level agent via OEM partnerships	ZTE Nubia collaboration; no self-manufacturing	Gen 2 phone in development with ZTE Nubia
OpenAI	Agent phone with app-less architecture	MediaTek chip plus Luxshare manufacturing	Mass production targeted first half 2027, per Kuo
Apple	Apple Intelligence, on-device model processing	Full proprietary silicon and hardware stack	Active; phased agent feature rollout ongoing
Google	Gemini agents across Pixel and Android OEM fleet	Pixel hardware plus Android OS licensing	Pixel 10 ships with background AI agent features
Huawei / Xiaomi / Honor	Proprietary AI OS with full system integration	Self-owned hardware plus dominant offline channels	Scaling aggressively across distribution network

For all its model capability and large user base, ByteDance lacks what every other player in that table possesses: hardware sovereignty and a distribution channel that functions without a screen. Among the companies listed, it is the only major AI contender that depends entirely on partners for both the device and the shelf space.

The second-generation Doubao phone, now in development with ZTE’s Nubia and expected to arrive in the coming months, will be the first real test of whether its app-negotiation strategy has advanced since December 2025. If cooperation agreements hold and Doubao secures system-level licensing across a broader manufacturer set, the platform play becomes credible. If the super apps resist again, those new OS engineers solve the technical layer but leave the commercial problem exactly where the Nubia M153 left it.

AI

Grok Build’s Four-Hour CRM Test Puts Custom Dev Firms on Notice

Published

17 hours ago

May 27, 2026

Logan Pierce

Grok Build, xAI’s new agentic coding agent, entered early beta on May 14, 2026, carrying a striking opening claim: a custom CRM (Customer Relationship Management system) with Salesforce and HubSpot data import pipelines, a working front-end, and drag-and-drop reporting, prototyped in under four hours. No independent test has confirmed the timing; xAI attributes the figure to preliminary user testing and internal showcases.

But even skeptics should look past the speed number. Grok Build spawns up to eight parallel AI sub-agents, runs code on the developer’s own machine rather than xAI’s servers, and has entered a market where boutique software firms have built careers around exactly the CRM complexity it aims to compress into an afternoon.

The CRM in Four Hours

Grok Build is a terminal-based CLI (command-line interface, a text-driven programming environment) designed for professional software engineering and complex coding work. A developer opens it inside a repository, describes a task in plain language, and the tool reads the codebase, maps an execution plan, and waits for explicit approval before modifying a file. Every change surfaces as a clean diff, so engineers can inspect and reject individual edits before anything is committed to the project.

The CRM prototype that circulated among early testers included data import pipelines from Salesforce and HubSpot, a functional front-end, and drag-and-drop reporting capabilities. Four-hour completion is the figure xAI is working with; the company has not published a controlled, reproducible demonstration of the build time and acknowledges the product is in early beta with active feedback loops shaping the next iteration.

Four headline numbers frame where the tool sits right now:

$299/month at full list price, with a six-month introductory rate of $99/month, accessible to SuperGrok Heavy and X Premium Plus subscribers
256,000-token context window on the grok-build-0.1 model, enough to hold a large codebase in memory across a full session
8 parallel sub-agents maximum, each assigned to an isolated Git worktree to prevent conflicting changes
70.8% on SWE-Bench Verified, the industry-standard agentic coding benchmark, per xAI’s internal testing

Elon Musk, xAI’s founder and chief executive, personally called for beta testers on X on May 14 and posted tips for early users that combined for more than 1.6 million views. The company added a /feedback command inside the CLI so developers can send bug reports without leaving the terminal, and on May 25 released a Windows PowerShell installer, extending the tool to the operating system that still runs the majority of enterprise desktops. xAI’s early beta announcement on x.ai frames the release as an iteration loop, not a finished product, and invites developers to shape what ships next.

Grok Build xAI coding agent early beta builds custom CRM in four hours.

Eight Agents, One Worktree Each

The parallel sub-agent architecture is the genuine technical differentiator. Most AI coding agents work sequentially, one model processing a chain of reasoning one file at a time. Grok Build assigns a coordinator agent to break a complex project into subtasks, then deploys up to eight specialized sub-agents simultaneously. The important distinction from Claude Code’s multi-agent mode is isolation: Claude Code sub-agents share the primary workspace, while Grok Build places each sub-agent in its own Git worktree so parallel branches can experiment and modify files independently without overwriting each other. In a complex refactor touching multiple modules, one sub-agent might be rewriting the database access layer while another restructures the UI components, and neither blocks the other’s progress. According to xAI’s Grok Build CLI documentation, each child sub-agent runs with its own context window and merges results when complete.

Three other design choices matter for enterprise teams evaluating the tool. First, Grok Build is local-first: source code stays on the developer’s machine and is not transmitted to xAI’s servers during a session, which is a meaningful distinction for contractors under NDAs and teams in regulated industries such as healthcare and finance. Second, headless mode (activated with a -p flag) lets teams embed the agent in CI/CD (continuous integration and continuous delivery) pipelines with no interactive interface. Third, the tool reads AGENTS.md instruction files, MCP (Model Context Protocol) servers, plugins, hooks, and skills from the existing project folder, so developers migrating from Claude Code carry their tooling configurations without rebuilding from scratch.

On the model side, xAI routed earlier grok-code-fast-1 requests to the newer grok-build-0.1 as of May 15, a migration schedule that signals the company is building a dedicated coding architecture rather than wrapping a general-purpose model in a terminal interface. The grok-build-0.1 model prices at $0.20 per million input tokens and $1.50 per million output tokens via API, well below the going rates for comparable Claude Opus or Codex calls at equivalent task volume.

Grok Build vs. the Field

The market Grok Build is entering already has two well-established players. Anthropic’s Claude Code, which launched in May 2025, has become the primary growth engine inside a company now tracking toward a $30 billion annualized revenue run rate driven largely by enterprise coding tool adoption, up from $14 billion just two months prior according to Bloomberg. OpenAI’s Codex CLI has surpassed three million weekly active users. According to JetBrains’ January 2026 AI Pulse survey of more than 10,000 professional developers, 90 percent of developers now use at least one AI tool at work, making coding the single largest enterprise generative AI use category by spending.

Tool	Developer	SWE-Bench Verified	Parallel Architecture	Subscription Entry
Grok Build 0.1	xAI	70.8%	Up to 8 sub-agents in isolated Git worktrees	$99/month intro; $299/month full
Claude Code	Anthropic	87.6% (Claude Opus 4.7)	Sub-agents in shared workspace	From $100/month (team tier)
Codex CLI	OpenAI	88.7% (GPT-5.5)	Cloud parallel tasks	Included with ChatGPT plans

Sources: xAI pricing page; Anthropic Claude pricing; DigitalApplied benchmark analysis, May 2026; vendor figures.

The benchmark gap is seventeen percentage points on SWE-Bench Verified, separating Grok Build’s 70.8% from Codex CLI’s 88.7% and Claude Code’s underlying 87.6%. For teams running a coding agent on consequential production changes, that spread is a genuine risk factor. Against it, Grok Build’s API token pricing undercuts comparable Opus or GPT-5.5 calls substantially, and the isolated Git worktree architecture for parallel sub-agents has no direct equivalent in either competitor’s current build. GitHub Copilot, with 4.7 million paid subscribers, operates as an IDE-first suggestion layer rather than an autonomous terminal agent. Google’s Jules and Gemini Code Assist Enterprise serve different workflow niches. The agentic terminal CLI category is, for now, a three-horse race.

The Custom Software Market Behind the Demo

Mordor Intelligence’s custom software development market forecast puts the global market at $50.94 billion in 2026, growing at a 17.88% CAGR to reach $115.95 billion by 2031. Enterprise migration from packaged applications toward bespoke solutions is the primary stated growth driver, and coding agents accelerate that migration by compressing the cost of the bespoke build in the first place.

$50.94 billion – estimated global custom software development market size in 2026, growing at 17.88% CAGR through 2031 (Mordor Intelligence)
78% of multinational corporations use custom platforms for ERP, CRM, and process automation, per industry surveys
1.2 million developer roles estimated to remain unfilled in the US alone by 2026, a scarcity that has sustained premium billing rates for outsourced development
39% of enterprises cite difficulty sourcing AI and cloud-native development expertise, the category of skill coding agents are designed to automate

Custom development shops price their services around two things: developer hours and the complexity of integration work. A Salesforce-to-HubSpot data pipeline migration, exactly the kind of engagement the CRM demo targets, can anchor a boutique firm’s pipeline for months. If a coding agent scaffolds that work in four hours rather than four weeks, the billable hour count drops, and so does the negotiating logic behind a six-figure statement of work.

The pressure will not fall evenly across the market. Large consultancies such as Accenture, Infosys, and Tata Consultancy Services, which together hold roughly a quarter of the custom software development market by revenue, compete primarily on compliance depth, security architecture, and multi-year transformation programs. AI-generated scaffolding code creates new complexity for those firms to govern rather than less, and their engagement models may expand as clients ship more AI-assisted prototypes that then need enterprise hardening.

Mid-tier boutique firms face the inverse dynamic. Their competitive advantage has been the talent gap: enterprises unable to recruit qualified developers outsource CRM integrations, portal builds, and reporting dashboards to them. As coding agents compress the hours required for those engagements, the talent gap that underpinned the billing relationship narrows. Harvard Business Review’s April 2026 analysis described generative AI as dissolving the economic logic that made standardized enterprise software the only practical choice for most companies. The corollary for the custom dev market is that it is also dissolving part of the logic that made hiring a boutique firm the only practical alternative. Those firms have roughly 12 to 24 months to reposition around work that coding agents cannot yet handle credibly: domain-specific compliance mapping, adversarial security auditing, and the organizational change management that surrounds any enterprise software rollout.

What Narrows the Gap

The 17-point benchmark deficit is real, but it carries an expiry date that is visible from today’s vantage. Grok 5, xAI’s next flagship model, is reported to carry 6 trillion parameters and a 1.5 million token context window, with release expected before mid-2026. Once Grok 5 powers Grok Build, the parallel worktree architecture stays unchanged while the model underneath it becomes considerably more capable, and the benchmark story changes accordingly.

Arena Mode and the Self-Ranking Bet

The more forward-looking feature is Arena Mode, confirmed in xAI code traces in February 2026 but not yet live in the current beta. Rather than presenting a single solution, Arena Mode runs multiple agents against the same problem, scores their outputs automatically, and surfaces the best-ranked result before the developer reviews anything. All agent responses appear side by side with a usage tracker, ordered by score, before any human decision is required.

The underlying logic is probabilistic. A coding model running eight parallel attempts and selecting the best result will outperform its baseline benchmark score considerably, because it gets to discard the seven weaker runs. Competing tools get one attempt per task; Grok Build, once Arena Mode ships, gets eight. Whether the self-ranking algorithm is reliable enough to consistently identify the genuinely best output remains the unanswered question, but the structural argument for why it raises effective performance is straightforward.

Platform Risk and the Cursor Equation

Enterprise teams evaluating adoption today face a complication that goes beyond benchmark scores. xAI completed its merger with SpaceX in February 2026, and according to reports, SpaceX has disclosed a $60 billion option to acquire Cursor, the IDE-native coding editor with an estimated $2 billion ARR, exercisable roughly 30 days after Cursor’s planned June 12, 2026 IPO. If that acquisition closes, xAI’s developer tooling story could include two complementary products: Grok Build for terminal-native workflows and a Cursor integration for the IDE-first majority. That combination would cover nearly every point in the development workflow. A beta-stage tool at a company mid-merger, with a large acquisition option pending, is nonetheless a more complex platform commitment than adopting Claude Code or Codex CLI today.

For teams comfortable with that context, the introductory pricing creates a reasonable entry point. At $99 per month for six months, Grok Build can run alongside an existing coding agent for less than the cost of a single developer day, and parallel evaluation is how most engineering teams actually decide whether a new tool belongs in their workflow.

If Arena Mode ships and a Grok 5 upgrade follows before the introductory window closes, the benchmark deficit narrows substantially and the architectural lead on parallel worktree execution looks more durable. If neither arrives on schedule, the full-price renewal hits before the product has earned it, and early adopters face a repricing conversation. The CRM demo opened the door; what comes through it depends on the next 90 days of execution.