According to a report by The Information cited by ithome.com.tw, Apple is negotiating to acquire an AI chip startup "to support the compute needs of AI applications and model training" (E1). The same report says Apple's AI servers currently run on the company's own M2 Ultra chip, but that chip "is not fast enough," which forced the new version of Siri to run on Google Cloud, whose underlying hardware is built on Nvidia chips (E2).
Compounding the pressure, Apple reportedly tried a different fix earlier this year: running Google's Gemini large language model on its own internal servers as part of the Siri upgrade. According to money.udn.com, that internal test's computing speed "did not meet expectations," so Apple ultimately adopted Google Cloud's AI infrastructure instead (E11). Meanwhile, Apple's self-designed AI server chip, code-named Baltra, was originally expected to ship this year, but Bloomberg reported the timeline may now slip to 2029 (E3).
The evidence points to a chip that is behind schedule and behind on performance. Per ithome.com.tw, the M2 Ultra-based AI server setup was too slow to keep new Siri features running in-house, forcing a shift to Google Cloud's Nvidia-based infrastructure (E2). The successor chip, Baltra, was supposed to ship this year but has fallen behind schedule, with Bloomberg's reporting pointing to a possible 2029 launch (E3). TechNews, reporting the same detail independently, confirms that "Apple originally planned to launch its next-generation AI server chip, code-named Baltra, this year, but the plan has been delayed" (E16).
On spending, the evidence shows a stark contrast. According to Wall Street analyst reports cited by ithome.com.tw, Amazon, Google, Microsoft and Meta have together committed roughly $650 billion to building AI data centers, with Meta alone planning to spend $145 billion to $165 billion this year. By comparison, Apple's capital expenditure has stayed at around $14 billion (E8).
| Company/Group | AI Infrastructure Spend |
|---|---|
| Amazon + Google + Microsoft + Meta (combined) | ~$65 billion... $650 billion |
| Meta (this year alone) | $145–165 billion |
| Apple (capex) | ~$14 billion |
Apple's cash position is large relative to its historical deal sizes. As of March 28, Apple held $45.57 billion in cash and cash equivalents, according to money.udn.com (E13), a figure also reported by technews.tw (E17). Yet Apple's largest acquisition to date remains the 2014 purchase of Beats for $3 billion (E4), and even the foundation of its self-designed chip business traces back to the $280 million acquisition of PA Semi (E5). Both figures are far below the valuations now attached to leading AI chip startups, raising the question of whether Apple would need to break from its historical deal-size pattern.
| Deal | Value |
|---|---|
| Beats (2014, Apple's largest acquisition to date) | $3 billion |
| PA Semi (foundation of Apple's chip design business) | $280 million |
| Apple cash & equivalents (as of March 28) | $45.57 billion |
According to ithome.com.tw, Apple has reportedly approached multiple chip startups to gauge acquisition interest, and money.udn.com separately confirms Apple "has contacted several chip startups to explore whether they are open to being acquired, and has discussed potential deals with investment banks" (E10). Named valuations in the reporting include Cerebras at $41.6 billion, SambaNova at $11 billion, Groq (already acquired by Nvidia) at a onetime valuation of $6.9 billion, Tenstorrent at $2.6 billion, Microsoft-backed d-Matrix at $2 billion, and Nvidia rival Etched at $5 billion (E6).
| Startup | Valuation |
|---|---|
| Cerebras | $41.6 billion |
| SambaNova | $11 billion |
| Groq (acquired by Nvidia) | $6.9 billion (peak) |
| Etched | $5 billion |
| Tenstorrent | $2.6 billion |
| d-Matrix (Microsoft-backed) | $2 billion |
One name stands out for its Apple ties: Tenstorrent's CEO is Jim Keller, a former Apple chip executive who led development of the A4 and A5 chips, giving him deep historical ties to the company (E7).
Apple is "seeking to acquire chip companies to strengthen its ability to develop AI server processors," according to money.udn.com (E9), a claim republished with identical wording by technews.tw (E15). This marks a departure from Apple's typically low M&A activity: its most recent deal was the January acquisition of Israeli AI startup Q.ai (E12), a detail also confirmed by technews.tw (E17). Separately from acquisitions, Apple has also been building supply partnerships — last week it announced a multi-year chip supply agreement with Broadcom valued at more than $30 billion, expanding its US-based chip procurement (E14).
Taken together, the evidence shows a company under performance pressure on two fronts: its current AI server chip (M2 Ultra) was too slow to keep Siri running in-house (E2), and its next-generation chip, Baltra, has slipped from a planned this-year launch toward a possible 2029 date (E3, E16). That timeline gap sits against a spending backdrop where rivals have committed hundreds of billions of dollars to AI infrastructure while Apple's capex remains near $14 billion (E8) — even though Apple's $45.57 billion cash pile (E13, E17) dwarfs its historic deal sizes of $3 billion for Beats (E4) and $280 million for PA Semi (E5). Whether an acquisition of a startup valued anywhere from $2 billion to $41.6 billion (E6) would close that gap, or simply buy time alongside the new $30-billion-plus Broadcom supply deal (E14), is not addressed in the available reporting.
As of 22 July 2026, standard API pricing (input/output per 1M tokens) and Artificial Analysis capability scores for major LLMs: OpenAI GPT-5.6 Sol ($5/$30, 59) and Anthropic Claude Opus 4.8 ($5/$25, 56) lead on capability; xAI Grok 4.5 and Claude Sonnet 5 sit in the value band; the DeepSeek V4 family prices output under $1 for high-volume use. Pricing changes often — verify on official pages.
According to a report by CNA (Central News Agency), the U.S. Commerce Department confirmed that Chris Fall, director of the Center for AI Standards and Innovation (CAISI), resigned after only three months in the role. The department did not disclose a reason, and a UDN report notes the departure marks the latest personnel shift in the Trump administration's AI policy.
According to Inside.com.tw, Taiwan's Legislative Yuan Transportation Committee has passed a preliminary review of an amendment to the Telecommunications Management Act that would exempt satellite operators from foreign-ownership caps under national-security review. Per CNYES, Chunghwa Telecom (中華電信) chairman Jie Zhicheng said the company would immediately move to bring in Starlink once the law passes, while Digital Affairs Minister Lin Yi-ching noted Starlink has told the government Taiwan's market potential is limited.