When "Jurassic Park" Goes to China

I spent last week in the French Polynesian islands for a quick vacation.  One of my favorite moments there was approaching the magnificent landscape of Moorea on a ferry, because it reminded me of one of my favorite movies growing up -- Jurassic Park.

My mind wanders in weird ways, especially when I’m off the grid. In this case, thinking about dinosaurs led my neurons to the semiconductor design technology that was used to render Jurassic Park, MIPS (Microprocessor without Interlocked Pipelined Stages), and an overlooked article that exposed MIPS as the latest American technology that has now become a key ingredient in China's pursuit of technological self-reliance.

How did MIPS end up in China

A little more than a month ago, Reuters published a detailed account of how MIPS got licensed to the Chinese company, CIP United, through a series of acquisitions, spin-offs, investments, and a bankruptcy. But the story did not get nearly as much attention as it deserved, publishing at a time when Huawei just got slapped with another round of sanctions, and the "fear and loathing" of TikTok and WeChat was approaching a new height -- both much shinier objects.

The entire story is worth reading to appreciate the complexity of the wheeling, dealing, and off-shoring that were necessary to on-shore a piece of important American technology to China. As for the tl;dr to grasp the transaction's current business and geopolitical significance, these are what I believe to be the most salient details:

  • CIP now controls the licensing rights of MIPS in China, Hong Kong, and Macau, a deal worth $60 million USD; one of its current customers is Huawei.
  • CIP has the right to develop derivative technologies based on MIPS. (This is not unusual for a proper license holder; companies like Broadcom and NEC have licensed MIPS to do the same.)
  • A 2019 CIP investor presentation explicitly stated its intent on using MIPS to help China fulfill its "Made in China 2025" plan.
  • CFIUS did not review this transaction, because it typically reviews direct investments, acquisitions, or joint-ventures, not licensing agreements.
  • Wave Computing, the company that currently owns the MIPS technology, filed for bankruptcy in April 2020, but its licensing agreement with CIP remains in effect.

The MIPS Journey

This technology transfer is significant, because the MIPS architecture is not an obscure research project or unproven tech from a young startup. It was once prominent in the semiconductor industry -- first developed by Stanford professor (later president) John Hennessey and his computer science students back in 1981. Its creation led to the founding of MIPS Computer Systems, which IPO’ed in 1989, then was acquired by Silicon Graphics and rebranded to MIPS Technologies, which was spun out and IPO’ed again in 1998, before declining as a company. During its prime, MIPS was considered a legitimate contender in the semiconductor design space among Intel and Arm. Besides rendering Jurassic Park movies, it was widely used to power gaming consoles like Nintento 64 and PlayStation 2, as well as many PCs and workstations during the 90s.

Eventually, its chip-related patents were sold to the UK company, Imagination Technologies, which a Chinese government-funded private equity firm called Canyon Bridge tried to acquire in September 2017. However, the MIPS portion of Imagination was forced to be spun out again, because not doing so would have triggered a CFIUS review -- a death knell particularly for Canyon Bridge, who was already on the CFIUS naughty list for its attempt to buy Lattice Semiconductor. As his first CFIUS review as president, Trump blocked the Lattice acquisition in that same September.

The wayward MIPS journeyed through a VC firm, a research lab, and eventually to Wave Computing, all of which are affiliated with the famed Filipino entrepreneur, Dado Banatao. The now-bankrupt Wave also counts the same Canyon Bridge plus Alibaba as minority investors (“minority” as in less than 10% ownership).

The Reuters team put together this graphic that condensed many of the important events into a helpful timeline:

Source: https://graphics.reuters.com/USA-CHINA/TECH/yzdvxxdlnvx/USA-CHINA-TECH.jpg

Although MIPS’s business traction is waning, it is still central to many university computer science programs when teaching the fundamentals of how a computer works. It was sadly not part of one of Stanford's foundational courses, CS107, that I took as a graduate student there, so I never had a chance to play with it.

MIPS’s Technology and China Implication

As a semiconductor design technology, MIPS falls in the RISC category, as opposed to CISC; it was in fact one of the first RISC-based chip architectures that gained wide usage.

The “R” in RISC stands for “reduced”, while the “C” in CISC stands for “complex”. I wrote a short technical explainer of RISC vs. CISC in a previous post “RISC-V, China, Nightingales”, so I won’t repeat those points here. There are many more detailed descriptions of these two instruction set architectures (ISAs) around the Internet, if you want to get into the weeds of it all. I won’t summarize them here, since it’s not that relevant to our discussion of MIPS and China.

One technical characteristic of MIPS that is relevant and worth noting is its design decision to move the software layer towards the hardware layer, as opposed to the reverse direction, which was the mainstream design philosophy at that time. This choice was at least in part due to the Hennessey-led research team’s strength in compilers, which is the layer of software that translates source code (i.e. your if-then statements) into machine code that a microprocessor can...well...process!

This is significant because making a MIPS-based chip perform well could mean needing more software than hardware expertise. This dynamic is likely true for all RISC-based chip architectures, because RISC is by definition a more “reduced” or simple set of instructions. Overall, there is more software than hardware talent and that trend will continue if only because software engineers are much better compensated than hardware ones at large. One place that is producing software engineers more quickly than anywhere else is China. The number of open source developers on GitHub, the dominant open source software collaboration platform, is a decent proxy for the total number of developers. China (minus Hong Kong) is now a close second to the U.S., while third place India still lags behind by a large margin. If you include Hong Kong, one of the fastest growing regions on GitHub, with China, the total number may in fact exceed that of the U.S.

Of course, not everything that happens on GitHub is strictly about software. The open source ISA, RISC-V, maintains an active set of code repositories. This software-over-hardware dynamic also does not mean China can catch up to the U.S. overnight simply with access to MIPS. But this design decision, made almost 40 years ago, may tip the scale for China in the long-run.

CFIUS Fail

While there are many more intrigues to the MIPS technology that we can dig into, its journey to China illuminates a larger issue: CFIUS is failing as a regulatory body. This Committee is ill-equipped and ill-designed to strike the balance between protecting national security, while maintaining a business environment that’s still open to foreign investments and not thwart economic competitiveness.

First created in the 1970s under President Gerald Ford, CFIUS hardly met at all during its first four decades of existence, thus has barely accumulated any institutional knowledge or domain expertise. Trump’s blocking of Canyon Bridge’s acquisition of Lattice was only the fourth time such an intervention has happened in CFIUS’s history up to that point. I’m not suggesting that more intervention is better, but the entire process still feels ad-hoc and politically-motivated, and thus lacks procedural justice and Due Process, which became more evident during the whole TikTok ordeal.

With the passing of the Foreign Investment Risk Review Modernization Act (FIRRMA) in August 2018, some CFIUS reform is under way. For one, the Treasury Department bureau in charge of leading CFIUS, the Office of Investment Security, has been bumped up to the Assistant Secretary level from the Deputy Assistant Secretary level, though the Assistant Secretary position was not filled until September 2019. For readers who aren’t familiar with the organizational structure of a Federal agency, if the Treasury Department was a company and the Treasury Secretary was its CEO, an Assistant Secretary is more like a VP and a Deputy Assistant Secretary is more like a Director. Effectively, the post-FIRRMA CFIUS moved up from middle management to upper-middle management -- more important but still not that important. FIRRMA also trains the CFIUS fire more squarely on China, without mentioning the country by name, and authorized the possibility of a “CFIUS fund”, so the Committee may, one day, have more money to hire more people and build more capacity to properly review more transactions, like the licensing agreement of MIPS to CIP.

These reforms always take a long time to materialize, and the Trump administration has been uniquely bad at attracting and retaining capable people to work in government. As CFIUS figures itself out, slowly and hopefully surely, technology transfer from America to elsewhere is only getting more complex -- the MIPS licensing not only meandered through a myriad of acquisitions and divestments, but also various entities in the Cayman Islands and Samoa.

At the time of this writing, there are six job openings on usajobs.gov when you search for CFIUS. It’s going to take more than that to build a regulatory body serious and capable enough to do its job.


If you like what you've read, please SUBSCRIBE to the Interconnected email list. To read all previous posts, please check out the Archive section. New content will be delivered to your inbox (twice per week). Follow and interact with me on: Twitter, LinkedIn.


侏罗纪公园去中国

我上周在法属波利尼西亚群岛过了个短暂的假期。让我印象最深的一刻是在渡船上接近摩尔岛的壮观山景,让我想起了小时候最喜欢的电影之一,《侏罗纪公园》。

我的脑袋经常会往奇妙的方向游荡,尤其是我在休息,没有被网络干扰的时候。这次我在想恐龙的时候就莫名其妙地游荡到了一款曾经用于数字化现实《侏罗纪公园》里各种图像的半导体设计技术,MIPS (Microprocessor without Interlocked Pipelined Stages),还有一篇被忽略了的文章揭露了MIPS做为一项美国技术,是怎么去的中国从而变成了中国追求科技自立自主的关键要素。

MIPS怎么去的中国

一个多月前,路透社发表了一篇详细的报道,讲述了MIPS是如何通过一系列收购、分拆、投资和破产使得一家中国公司,CIP United,得到它的技术许可。但这篇报道并没有得到应有的关注,它发表在华为刚刚受到新一轮制裁的时候,对TikTok和微信的“恐惧和厌恶”也正在接近一个新高潮——两个都是更大的“亮点”。

整个文章都值得大家一读,以了解其中怎么把美国技术“运往”中国所必需做的各种转移、交易和离岸业务的复杂性。从了解整个事件目前对商业和地缘政治影响的角度来看,我觉得以下的细节最值得关注:

  • CIP现在控制MIPS在中国、香港和澳门的独家许可权,是一笔价值6000万美元的交易;目前的一大客户之一是华为。
  • CIP有权开发基于MIPS的衍生技术。(这对一个有正当许可的公司来说并不奇怪;像Broadcom和NEC这些大公司都曾经买过MIP的许可证做同样的事情。)
  • 一份2019年CIP面对投资人的报告里明确表示,其意图是利用MIPS帮助中国实现其“中国制造2025”计划。
  • CFIUS没有审查这项交易,因为它的通常审查范围是直接投资、收购或合资企业,不是许可协议。
  • 目前拥有MIPS技术的Wave Computing于2020年4月申请破产,但其与CIP的许可协议仍然有效。

MIPS之旅

这项技术的转让意义重大,因为MIPS不是一个不起眼的学术科研项目,也不是一个年轻的创业公司还未经验证的技术。它曾经在半导体行业里的佼佼者之一,在1981年由斯坦福大学教授(后来的校长)John Hennessey和他的计算机系学生开发的。随后是MIPS Computer Systems的成立,该公司于1989年首次上市,随后被Silicon Graphics收购,并更名为MIPS Technologies,后者于1998年再次上市,随后开始衰落。在其鼎盛时期的90年代中,MIPS被认为是英特尔和Arm在半导体设计领域的竞争对手。除了呈现《侏罗纪公园》以外,基于它的半导体还被广泛应用在像Ninento 64和PlayStation2的游戏机里,以及许多PC和大型工作站。

最终,其芯片相关的专利被出售给一家英国公司Imagination Technologies。一家有中国政府出资的PE资本公司Canyon Bridge曾经在2017年9月试图收购Imagination。但MIPS这一块被迫再次被分开,因为如果不这样做会引发CFIUS审查,对Canyon Bridge来说没有好结果,因为Canyon Bridge已经上了CFIUS的“黑名单”。这家基金曾企图收购另外一家美国半导体公司,Lattice Semiconductor,但最终被特朗普否决。这项否决是特朗普做为总统的第一次CFIUS审查,也发生在同一个9月。

随后,继续流荡的MIPS去了一家风投,一个研究实验室,最终到了Wave Computing,而这几个组织都于菲律宾裔企业家Dado Banatao有关系。如今破产的Wave公司里的两个少数股东(“少数”定义为持股10%以下)测试同一家Canyon Bridge和阿里巴巴。

路透社团队做了这幅非常好用的图表,总结了许多这些重要事件放在了一个时间轴上:

Source: https://graphics.reuters.com/USA-CHINA/TECH/yzdvxxdlnvx/USA-CHINA-TECH.jpg

尽管MIPS的商业应用今非昔比,但在各大计算机系的基础课程里做为教育素材,仍然被广泛使用。可惜我在斯坦福当研究生时,老师没有用MIPS教我拿的其中一门基础课(CS107),所以我个人还没有机会玩过它。

MIPS这门技术与它对中国的意义

作为一款半导体设计技术,MIPS属于RISC派系,而不是CISC;它实际上是最早获得广泛应用的基于RISC的芯片体系结构之一。

RISC中的“R”代表“reduced”,CISC中的“C”代表“complex”。我在之前的一篇文章《RISC-V,中国,夜莺》中写了一段关于RISC与CISC的简短技术说明,在这里就不重复了。如果您想深入了解这两种硬件指令集体系结构(Instruction Set Architecture,ISA),网上有很多更详细的资料。我在这里就不一一总结了,因为与我们讨论MIPS和中国科技的“互联”没有太大关系。

但MIPS内部的一项技术特征是非常相关且值得注意的,那就是它的核心设计理念是使软件层靠向硬件层,而不是相反的方向(相反方向是当时的主流设计理念)。这一选择至少在一定程度上是由于Hennessey领导的研究团队在编译器方面很强,编译器则是将源代码(比如一条if-then语句)翻译成芯片看的懂得机器代码来处理。

这点很重要,因为要把基于MIPS造的芯片的性能做的好,可能需要更多的软件而不是硬件知识。这一特性也可以概括到所有基于RISC的芯片架构,因为RISC的定义就是“reduced(精简)”或“简单”的一套指令集。总的来说,软件人才多于硬件人才,这一趋势也会持续下去。简单归因,总体来说软件工程师比硬件工程师要赚的多得多。中国则是软件工程师培养速度最快的地方之一。用在GitHub(主流开源软件协作平台)上的开源开发者的数量一个较好的估算所有开发者总数的代表。中国的开发者数量(不包括香港)现在仅次于美国,而位居第三的印度还落后一大截。如果把香港加进去(香港是GitHub用户增长最快的地区之一),拿中国的总数很可能已经超过了美国。

当然,GitHub上的各种开发并不都是软件。RISC-V(一项开源的ISA项目)就在GitHub上有一组很活跃的代码库。这种软件胜过硬件的趋势也并不意味着只要有MIPS,中国就会在一夜之间赶超美国。但从长远来看,这项40年前做出的技术设计决定,很可能对中国科技发展的未来有深远的影响。

CFIUS的失责

MIPS的故事和技术还有很多我们可以聊的细节,但整个事件和它的中国之旅揭示了一个更大的问题:CFIUS作为一个监管机构是失责的。这个委员会的整个设计和能力目前都无法有效的保护美国国家安全,同时维护一个正常的商业环境和对国外投资的开放姿态,以不阻碍美国的经济竞争力和发展。

CFIUS是70年代在福特总统的领导下创立的。委员会成立的头40年里几乎没有见过面,因此也没有积累任何能力或专长。特朗普否决Canyon Bridge收购Lattice,仅是CFIUS历史上第四次发生此类的干涉。我并不是说干涉越多越好,但整个过程仍然让人觉得很临时而武断,以及有政治动机,从而缺乏程序正义以及正当程序,这一点在TikTok事件上变得更加明显。

随着2018年8月《外国投资风险审查现代化法案》(The Foreign Investment Risk Review Modernization Act,FIRRMA)的通过,CFIUS正在进行某些改革。其一,负责领导CFIUS的财政部局投资安全办公室(Office of Investment Security)已从副助理部长(Deputy Assistant Secretary) 级晋升为助理部长级 (Assistant Secretary),不过这一新的助理部长职位直到2019年9月才被填补上。对于不熟悉联邦机构组织结构的读者来说,我来打个比方:如果财政部是家公司,财政部长是公司的CEO,那助理部长更像是副总裁(VP),副助理部长更像是个Director。所以FIRRMA通过后的“新CFIUS”被“提拔”了,从中层管理层升到了中高管理层。它更重要了,但其实也并不是那么重要。FIRRMA还把CFIUS的矛头更明确的指向了中国,虽然没有直接提中国的名字,并授权设立一个“CFIUS基金”的可能性,因此,也许有一天,委员会会有资金雇更多的人员,打造更多的专业能力来合理正确的审查更多的交易,比如MIPS与CIP的许可协议。

这类改革总是需要很长的时间才能实现,而特朗普政府在吸引和留住有能力的人才到政府工作这一点是出了名的失责。在CFIUS还在”发掘自己“的同时,从美国向他国的各种技术转移只会变得越来越复杂——MIPS去中国的许可不仅通过无数项收购和撤资,还包括开曼群岛和萨摩亚的各种实体。

在写本文时,在usajobs.gov网站上仅有六个与CFIUS有关的职位空缺。要想建立一个严肃、有能力的监管机构来履行它的职责,这远远不够。

如果您喜欢所读的内容,请用email订阅加入“互联”。要想读以前的文章,请查阅《互联档案》。每周两次,新的文章将会直接送达您的邮箱。请在TwitterLinkedIn上给个follow,与我交流互动!