• Mon - Fri 9.00 - 19.00
Open

Tag Archives: China

Tamara's World  /  Blog  /  China
26 Sep 2026
Trump Xi summit trade White House Washington

Trump and Xi Bought Themselves Until January — But Not Much Else

Categories Business, World

The Trump Xi summit trade truce reached at the White House this week extends an existing tariff ceasefire through January 10, 2027. It buys both governments more time to negotiate a broader deal rather than resolving the underlying dispute. The September 24-26 meeting was the third in-person encounter between the two leaders in under a year, and it produced an extension rather than a breakthrough.

What the Trump Xi summit trade truce covers

The original 11-month truce had been set to expire in November. According to Al Jazeera’s summit preview, the dispute traces back to Trump’s 10% duties on Chinese goods, imposed over fentanyl and immigration concerns, which China answered with levies on US coal, LNG, crude oil and autos, plus restrictions on rare-earth metal exports. Trump later added 12.5% tariffs on 60 trading partners, including China, over forced-labor allegations, widening the dispute beyond a purely bilateral fight.

Taiwan arms sales remain the sharpest edge

Trump Xi summit trade shipping containers at port

Taiwan was expected to come up prominently, and it did. The Trump administration approved an $11 billion arms package for Taiwan in December, with a second $14 billion package currently on hold. Trump has described the sales as “a very good negotiating chip,” language that signals Washington sees Taiwan-related leverage as a bargaining tool in the broader trade relationship rather than a fixed commitment.

AI competition adds a new front

Beyond trade and Taiwan, the two sides discussed what officials describe as “fierce competition” in artificial intelligence development. The US has proposed a new notification mechanism under which each government would alert the other to serious AI incidents that threaten national security. It is an early and narrow step toward managing AI risk between the two largest economies, not a comprehensive framework.

Iran hovered in the background

Given the US-Israel conflict with Iran, that issue was expected to feature prominently in the talks as well. China maintains close ties with Tehran, and some observers suggested Washington might look to Beijing for influence over Iran’s posture. Trump, however, has indicated he does not believe such intervention is necessary, suggesting the Iran track stayed separate from the trade and Taiwan discussions rather than becoming a bargaining chip itself.

Coverage from NPR described the two leaders striking a cordial public tone throughout the three days. Officials on both sides still acknowledged that underlying tensions on trade, technology and Taiwan remained largely where they started.

How markets read the outcome

Equity markets took the extension calmly rather than treating it as either a major relief or a fresh shock. That fits a summit that avoided both a collapse in talks and a comprehensive new agreement. Analysts covering US-China trade said the January deadline gives negotiators a defined runway. They cautioned that previous extensions this year had similarly kicked difficult decisions further down the calendar rather than resolving them.

Business groups on both sides of the Pacific welcomed the extension as preferable to a lapse in the truce. Even so, several trade associations noted that repeated short-term extensions make it harder for manufacturers and exporters to plan around a stable long-term tariff schedule.

The clock now running to January

The truce is now extended to January 10, 2027. That gives both governments roughly three and a half months to negotiate terms on tariffs, rare-earth exports and the still-pending second Taiwan arms package. Analysts following the relationship say the extension avoids an immediate escalation but leaves every major flashpoint, tariffs, Taiwan, and AI, unresolved heading into the new deadline.

More on US-China trade

Related coverage: the Federal Reserve’s latest rate decision, AMD’s trillion-dollar valuation, and Jensen Huang’s comments on AI regulation.

Summit FAQ: trade truce explained

How long does the new trade truce last?

It runs through January 10, 2027, extending the original 11-month ceasefire that had been due to expire in November.

What started the tariff dispute?

Trump’s 10% duties on Chinese goods over fentanyl and immigration concerns, which China answered with levies on US coal, LNG, crude oil and autos plus rare-earth export restrictions.

What is happening with Taiwan arms sales?

An $11 billion package was approved in December; a second $14 billion package remains on hold as Trump calls the sales a negotiating tool.

Did the summit address AI competition?

Yes. The two sides discussed a proposed notification mechanism to alert each other to serious AI incidents threatening national security.

Was Iran part of the summit agenda?

It was expected to come up given the US-Israel conflict with Iran and China’s ties to Tehran, though Trump indicated he sees no need for Chinese intervention.

Summit reporting sources

21 Sep 2026
Data center hardware representing the StepFun Step 5 Preview AI model launch

A Chinese Startup Just Undercut OpenAI on Price by a Wide Margin

Categories Tech

Chinese AI lab StepFun launched its flagship model on September 20. The StepFun Step 5 Preview is already turning heads for one reason as much as any benchmark: price.

The 600-billion-parameter sparse mixture-of-experts model went live with API pricing of $1.00 per million input tokens. Output runs $2.70 per million tokens. That undercuts several Western frontier labs by a wide margin. It still scores competitively on independent benchmarks.

Server room hardware powering the StepFun Step 5 Preview AI model

What the StepFun Step 5 Preview Actually Is

Step 5 Preview carries roughly 600 billion total parameters. About 27 billion activate per token. That sparse architecture keeps inference costs down while preserving capacity. It supports a one-million-token context window. It takes both text and image inputs. The design targets long, complex tasks rather than short chat exchanges.

How It Stacks Up Against the Field

Step 5 scored 44 on the widely tracked Intelligence Index. That matches Kimi K3 Max, another Chinese frontier model. In one benchmark test, it completed a 24-hour autonomous optimization task. It ran at 508 TFlops on H100 GPU cores. StepFun points to that result as evidence the model can sustain long, agentic workloads without human help.

Matching a well-regarded rival like Kimi K3 Max does not by itself prove Step 5 leads the field. But it places StepFun firmly among the top tier of Chinese labs. Those labs now compete directly with the largest Western AI companies on capability, not just cost.

The Pricing Move That Has Rivals Watching

StepFun built the release around long-horizon agent workloads. Think AI coding, software engineering, financial analysis and professional knowledge work. Pairing that focus with aggressive pricing puts direct pressure on Western labs. Enterprise customers running high-volume agentic tasks watch token costs closely. Those costs add up fast at scale.

Open Weights Land October 15

Unlike many frontier releases that stay fully closed, StepFun has committed to opening the model’s weights on October 15, 2026. That gives outside researchers and companies a firm date to plan around. It continues a pattern of Chinese labs using open weights as a lever against closed Western rivals.

Why Chinese Labs Keep Undercutting on Price

StepFun’s pricing move follows a broader pattern among Chinese AI labs this year. Several have used aggressive token pricing to win developer mindshare. That holds even when their raw benchmark scores trail the very top Western models. The strategy leans on a simple bet. Once developers build workflows around a cheaper model, switching costs make it harder for pricier rivals to win them back.

That approach has drawn scrutiny from analysts. They question whether the pricing holds up at scale. For now, though, it has pulled models like Step 5 into serious enterprise evaluations. Those evaluations might otherwise have gone straight to established Western providers.

Where StepFun Fits Among Chinese AI Labs

StepFun is one of several Chinese AI labs that have released a series of models under a numbered naming scheme, steadily scaling parameter counts with each generation. Step 5 Preview is the company’s most ambitious release to date, both in size and in the workloads it targets.

Unlike rivals that lead with chat or search products, StepFun has positioned this release around agentic and enterprise workloads from the start. That focus shapes how the company talks about the model. Its own materials emphasize coding and financial-analysis benchmarks over general conversational ability.

What This Means for Enterprise AI Buyers

Enterprise teams evaluating AI vendors now have another low-cost option to weigh alongside the established players. Procurement teams say token pricing has become a bigger factor in vendor selection as usage scales into the billions of tokens monthly. A cheaper model that performs adequately can save real money at that volume, even if it lags top benchmarks slightly.

Data residency and compliance concerns may still steer some Western enterprise buyers away from Chinese-hosted models, regardless of price. That question is likely to shape how far StepFun’s pricing advantage actually translates into enterprise contracts outside China.

What to Watch Before the Weights Drop

The real test comes once independent developers get hands-on access to the open weights next month. Benchmark scores from a lab’s own team invite scrutiny. The coming weeks will show whether Step 5’s real-world coding and agent performance holds up outside StepFun’s own testing environment.

Until then, the API launch already lets outside developers test the model directly. That should produce a clearer, independent picture of its performance well before the October 15 open-weights date.

Straight Answers on StepFun’s New Model

What is the StepFun Step 5 Preview?

A 600-billion-parameter sparse mixture-of-experts AI model from Chinese lab StepFun, launched September 20, 2026, with about 27 billion parameters active per token.

How much does it cost to use?

API pricing is $1.00 per million input tokens and $2.70 per million output tokens.

What is it designed for?

Long-horizon agent workloads such as AI coding, software engineering, financial analysis and professional knowledge work.

How big is its context window?

One million tokens, with support for both text and image inputs.

When do open weights arrive?

StepFun has scheduled the model’s weights to open on October 15, 2026.

Sources

More AI coverage on Tamara News: OpenAI just admitted its models are lying and hiding mistakes and Perplexity’s new AI agent wants to run your PC for you.

21 Sep 2026
Semiconductor fab floor tied to CXMT G5 DRAM production in China

China’s Chip Champion Just Closed the Gap on Samsung and SK Hynix

Categories Tech

CXMT G5 DRAM production is now underway in China. It marks one of the clearest signs yet that the country’s chipmakers are closing the gap with Samsung and SK Hynix. That progress comes despite years of export restrictions on advanced lithography gear.

Chinese memory chipmaker CXMT announced the milestone at the 2026 World Manufacturing Convention in Hefei on September 20. It unveiled its fifth-generation DRAM platform, known as the G5 Platform.

Close-up of a memory chip representing CXMT G5 DRAM production

What CXMT G5 DRAM Production Actually Achieves

The G5 Platform uses quadruple patterning. That reaches an 11.95-nanometer active-area half-pitch for memory arrays. CXMT said its capacitor aspect ratio reaches 45:1. The core cell array height has shrunk to 6,762 nanometers. In plain terms, the company packs more memory into a smaller footprint. It says the platform yields at least 50% more dies per wafer than its predecessor.

How CXMT Got Here Without ASML’s Banned Machines

The achievement stands out for one reason. CXMT built it without access to ASML’s most advanced extreme ultraviolet lithography tools. Those remain subject to export controls. The company instead relied on quadruple patterning with existing equipment. That workaround took years to refine. It appears to be paying off at scale now.

Multiple patterning is not new in chipmaking. But matching the density gains EUV lithography delivers in a single pass usually costs more. It means more process steps, more time and more cost per wafer. CXMT chose to absorb that complexity rather than wait for policy changes. That choice reflects the broader bet Chinese chipmakers have made since export controls tightened.

What This Means for Samsung, SK Hynix and Global Memory Prices

CXMT now claims the title of China’s largest DRAM supplier. It also claims fourth place worldwide by market share. The company has scaled fast. It grew from roughly 40,000 wafers a month in 2020 to about 300,000 today. That output spans its 12-inch fabs in Hefei and Beijing. It targets 350,000 to 375,000 wafers monthly by year-end 2026. That capacity growth adds real supply to a global memory market squeezed by AI-driven demand.

The Products Already Shipping on the New Platform

Two 24-gigabit LPDDR5X products are already in mass production. They come in 496-ball and 245-ball packages for smartphones and portable electronics. Both are integrated into mainstream Chinese flagship handsets already, the company says.

The Bigger Picture for China’s Chip Ambitions

CXMT’s progress fits a wider Chinese push to build a self-sufficient semiconductor supply chain. That priority hardened after successive rounds of US export controls on advanced chipmaking equipment. Beijing has funneled heavy state-backed investment into domestic memory and logic chip production for years. CXMT’s rise from a minor player to a top-four global DRAM supplier is one of the clearest results so far.

Whether the company can keep closing the gap at the cutting edge remains an open question. Both Korean rivals keep investing heavily in their own next-generation nodes. Neither shows signs of slowing down.

Which Chinese Phone Makers Are Buying In

CXMT has not publicly named every customer using the new LPDDR5X chips. Industry watchers tracking teardown reports say several major Chinese smartphone brands have already qualified the parts for mainstream, non-flagship models. That matters commercially. Winning approval in China’s huge domestic phone market gives CXMT steady volume, even before it competes for premium international designs.

Analysts say the bigger prize is qualification with global device makers outside China. That would require CXMT to match not just density but also long-term reliability data that Samsung and SK Hynix have built up over decades. CXMT has not said when, or whether, it plans to pursue that qualification process.

Where CXMT Goes From Here

Capacity is still climbing toward its year-end target. CXMT’s next test is sustaining yields at scale while narrowing the technology gap further. Industry watchers expect the company’s progress to keep pressuring global memory pricing. That dynamic is already squeezing device makers worldwide this year.

The next milestone to watch is server and data-center memory. Can CXMT push the G5 platform into that higher-capacity segment? It remains dominated even more heavily by Samsung, SK Hynix and Micron than the mobile market CXMT has targeted so far. Analysts see that as the real test of whether CXMT can compete at the very top of the market, not just at the budget end.

Quick Answers on China’s New Memory Chip

What is CXMT G5 DRAM production?

It is the mass production of CXMT’s fifth-generation DRAM platform, announced at the 2026 World Manufacturing Convention in Hefei on September 20.

How advanced is the new chip?

The G5 Platform achieves an 11.95-nanometer active-area half-pitch using quadruple patterning, without relying on ASML’s most advanced banned lithography machines.

What products use it already?

Two 24-gigabit LPDDR5X chips, packaged for smartphones and portable electronics, are already in mass production and integrated into mainstream Chinese flagship phones.

How much is CXMT producing?

The company has grown from roughly 40,000 wafers per month in 2020 to about 300,000 wafers per month across its Hefei and Beijing fabs, targeting 350,000 to 375,000 by year-end 2026.

Where does CXMT rank globally?

CXMT now claims to be China’s largest DRAM supplier and the world’s fourth-largest by market share.

Sources

Related coverage on Tamara News: your next phone costs more because AI bought the memory first and Nvidia’s $12.9B Hugging Face deal.

14 Sep 2026
Racks of servers in a data centre, illustrating the China intelligent computing plan

China Intelligent Computing Plan Targets 9,800 EFLOPS by 2030

Categories Tech

China’s Ministry of Industry and Information Technology has published a five-year blueprint that puts a number on the country’s AI ambitions. The China intelligent computing plan, set out in the 15th Five-Year Plan for the information and communications industry released on 7 September 2026, targets 9,800 exaflops of intelligent computing capacity by 2030 — roughly six times the 1,590 exaflops baseline it cites for 2025 — supported by 3.8 trillion yuan of cumulative information infrastructure investment over the period.

Sections

Reading the target honestly

These are the Chinese government’s own targets, published by the ministry that will be judged against them, and Chinese five-year plans have a mixed record of delivery. They are still worth reading, because they reveal what the state has decided to spend on and how it defines success.

Two features stand out. First, the metric is compute capacity rather than model capability — the plan commits to building the substrate, not to producing a particular system. Second, the scale of the increase is front-loaded on infrastructure: reporting on the plan pairs the 9,800 exaflops figure with the 3.8 trillion yuan investment envelope, which is on the order of $530bn at recent exchange rates.

A note on units, because the figures get quoted loosely. Intelligent computing capacity as Chinese planners define it is measured at lower numerical precision than the FP64 benchmarks used for traditional supercomputer rankings. It is not directly comparable to the top of the TOP500 list, and anyone treating the two as equivalent is comparing different things.

The constraint the plan is built around

US export controls have, since 2022, progressively restricted Chinese access to the highest-performing AI accelerators and, increasingly, to the equipment used to make them. A national compute plan drafted in 2026 therefore cannot assume access to Nvidia’s leading parts, or AMD’s, or Intel’s.

The plan is explicit about building on domestic supply. That has a practical consequence that is easy to miss: hitting a capacity target with less capable individual chips means using far more of them, which raises the bill for power, cooling, floor space and interconnect. Reporting by DigiTimes points to ambitions around very large clusters — in the range of 100,000 accelerator cards — which is consistent with that arithmetic. Energy, not silicon, becomes the binding constraint at that scale.

It also means the plan is a demand signal to China’s domestic chipmakers, foundries and equipment vendors as much as it is a compute target. Guaranteed state-backed demand is how you finance a supply chain that cannot yet compete on unit economics.

The software half of the strategy

Alongside the infrastructure plan, the ministry issued an implementation plan for an “Artificial Intelligence Plus Software” action, reported in the days that followed. Its stated aims include optimising intelligent programming toolchains, pushing enterprises toward technical transformation measured on code quality and R&D productivity, and embedding automated security testing across the software development lifecycle — with an ambition to take Chinese professional software to world markets by 2030.

The through-line is import substitution applied one layer up the stack. Compute is the expensive part, but design and development tools are the part where dependency is hardest to unwind, because it is embedded in engineers’ habits and in decades of existing designs. China’s broader innovation agenda for the 2026 to 2030 period has been outlined in similar terms by the science and technology ministry, whose plans were published on the State Council’s English portal.

What this changes outside China

For chipmakers outside China, the plan confirms what export controls already implied: the Chinese AI-accelerator market is closing as an addressable opportunity, and the competitive question shifts to third markets. For Western AI labs, a compute build-out of this scale narrows the assumption that access to frontier compute is a durable moat.

For energy markets, it is another large data-centre demand curve landing on a grid that has to be planned years in advance — the same pressure visible in Europe and North America, at a different scale and with different politics.

What to watch through 2027: whether reported capacity additions track the trajectory implied by a 2030 target, how much of the build uses domestic accelerators versus stockpiled or grey-market foreign parts, and whether power availability starts appearing as the published constraint rather than chips.

Questions

What does the China intelligent computing plan target?

9,800 exaflops of intelligent computing capacity by 2030, up from a stated 2025 baseline of 1,590 exaflops, backed by 3.8 trillion yuan of cumulative information infrastructure investment.

Who published it and when?

The Ministry of Industry and Information Technology, through its information and communications development department, published the 15th Five-Year Plan for the industry on 7 September 2026.

Is that comparable to supercomputer rankings?

No. Intelligent computing capacity is measured at lower numerical precision than the FP64 benchmarks used in traditional supercomputer lists, so the figures are not directly comparable.

Does the plan rely on Nvidia chips?

No. It is built around domestic supply, reflecting US export controls that restrict Chinese access to leading foreign AI accelerators and manufacturing equipment.

What is the AI Plus Software action?

A separate implementation plan from the same ministry covering intelligent programming toolchains, enterprise technical transformation and automated security testing, with an aim of exporting Chinese professional software by 2030.

Should these targets be treated as forecasts?

They are government objectives published by the ministry responsible for meeting them, not independent projections, and should be read with that interest in mind.

01 Sep 2026
DeepSeek funding round

Inside the DeepSeek Funding Round That Could Precede a 2027 IPO

Categories Tech

A DeepSeek funding round worth roughly 50 billion yuan, about $7.4 billion, is nearing a close. Sources cited by China Money Network put the pre-money valuation near $74 billion. DeepSeek is the Chinese AI lab that rattled US markets in January 2026. Its model matched Western systems at a fraction of the training cost. DeepSeek is using this raise to fund about 1 gigawatt of new compute capacity, ahead of a planned initial public offering.

What the DeepSeek funding round is actually paying for

The bulk of the new capital is earmarked for compute infrastructure, not research headcount. That reflects how much of the competitive pressure in frontier AI now comes from access to chips and data-center capacity, rather than algorithmic breakthroughs alone. DeepSeek is racing rivals including Alibaba’s Qwen, Tencent’s Hy4, and Zhipu’s GLM for both compute and engineering talent. Sources familiar with the round say the financing is intended to close by the end of August. That would position DeepSeek to file for an IPO before the end of 2026, with a public debut targeted for 2027 on Shanghai’s Star Market.

DeepSeek funding round

Image credit: Feliciagrace-bytesrack / Wikimedia Commons (CC BY 4.0)

Why this funding round matters beyond DeepSeek

DeepSeek’s January 2026 model release briefly wiped more than $500 billion from AI-linked US equities. It demonstrated that competitive AI systems no longer required the enormous training budgets US labs had assumed were necessary. This new funding round suggests the opposite lesson is now taking hold inside DeepSeek itself. Staying competitive at the frontier still requires massive, sustained capital, even for a lab that built its reputation on efficiency. That shift has implications for every AI lab claiming a cost advantage. It implies efficiency gains buy time rather than permanently lower capital requirements.

The chip-supply backdrop shaping the round

DeepSeek’s fundraising lands against a backdrop of tightening US export controls on advanced AI chips. The Commerce Department has signaled further regulatory action on chips and AI is coming. Chinese AI labs have had to route around restrictions on the most advanced Nvidia and AMD chips. A war chest of this size gives DeepSeek more room to secure whatever compute capacity remains available to it. That capacity could come through domestic chipmakers or through channels that comply with existing export rules.

Analysts tracking China’s AI sector say the round also signals confidence from domestic investors that Beijing will keep backing frontier AI development despite the external pressure. That confidence matters as much as the capital itself, since it shapes whether other Chinese labs can raise on similar terms.

How investors are pricing DeepSeek against its global rivals

A $74 billion valuation puts DeepSeek well below OpenAI and Anthropic on paper. It still ranks among the most valuable AI labs outside the United States. Private investors backing the round are betting that DeepSeek can keep its cost advantage even as it spends more on infrastructure. That bet carries real risk. Training costs across the industry have climbed as models grow larger. DeepSeek’s efficiency edge could narrow if rivals adopt similar techniques.

Domestic Chinese funds make up a large share of the round, according to people familiar with the terms. That matters for a specific reason. It reduces DeepSeek’s exposure to foreign capital controls and export-linked investment restrictions. Those restrictions have complicated fundraising for other Chinese tech companies this year.

Some investors involved in the round declined to comment publicly, citing the sensitivity of ongoing negotiations. That reticence is typical for pre-IPO rounds of this size, where terms often shift until the final signing. Bankers advising on the deal have also stayed quiet about the exact investor list. That silence is standard practice ahead of a Shanghai listing, where regulators scrutinize pre-IPO ownership disclosures closely and expect confidentiality until filings become public.

What happens next

Investors will be watching whether the round closes at the reported terms by DeepSeek’s end-of-August target. They will also watch whether an IPO filing follows on the timeline sources have described. A successful Shanghai listing would give Chinese investors direct exposure to one of the country’s most closely watched AI companies. It would also test whether public markets value DeepSeek’s efficiency reputation as highly as private investors currently do.

Frequently Asked Questions

How much is DeepSeek raising in this funding round?

Reports put the round at roughly 50 billion yuan, about $7.4 billion, at a pre-money valuation near $74 billion.

What will DeepSeek use the money for?

The company plans to use most of the capital to build roughly 1 gigawatt of new compute capacity, according to sources familiar with the round.

Is DeepSeek planning an IPO?

Yes. Sources say DeepSeek could file for an IPO by the end of 2026, with a public debut targeted for 2027 on Shanghai’s Star Market.

How does this relate to DeepSeek's January 2026 model release?

That release showed DeepSeek could match Western AI systems at a fraction of the training cost, briefly erasing over $500 billion from AI-linked US stocks. This funding round suggests staying competitive still requires large, sustained capital investment.

Related Coverage

Sources