集微访谈 | Eric Bouche：半导体产业复苏，有一个“雁阵”

集微网消息，在往期的集微访谈栏目中，爱集微有幸采访了Silicon Valley Research Institute（SVRI）创始人兼董事总经理Eric Bouche。集微访谈就关于行业发展、市场供需周期、行业人才短缺以及AI发展前景等一系列问题，收到了十分有启发的答复。

问：行业目前的主旋律是什么？

答：一个重要的主题是行业的复苏，在疫情期间，我们经历了短缺的情况。然而，在疫情后，我们观察到了半导体行业发展减速和供过于求的情况，现在我们终于开始从这一阶段中复苏起来。行业正在逐渐恢复增长。然而，使得这一复苏与众不同的是大量的政府干预和补贴。据我所知，各国政府提供的补贴总额已经达到了数万亿美元，这是相当惊人的数字。因此，预计这一复苏将伴随着许多偏离正常状态的情况。

问：您谈到半成品行业复苏，但行业状况如何？

答：这是一个很好的问题，当我们谈论（半导体）行业时，通常会将所有领域归到一起当作一个大行业讨论行业动态。但这么做并不准确。实际上，如果你看一下晶圆厂、逻辑芯片、DRAM、3D NAND闪存等方面，复苏的情况是各不相同的。所以我倾向于说我们正在复苏，但这些领域的复苏程度各不相同。我认为晶圆厂，可能还有逻辑芯片是前面的领头羊，通常情况就是这样。这样说是有充分的理由的。我认为DRAM和3D NAND在价格曲线的底部勉强趋于平稳。而且，我们的一些客户是存储制造商，他们对目前的情况还不满意。我们知道它迟早会恢复，但现在还没完全到达能称得上恢复的水平。

同时虽然我前面已经提到我们已经走出了供不应求的状态，但这并不完全正确。我知道总会有人告诉我我们仍然处于短缺状态。有一些领域仍然存在短缺，我想提到其中一个领域，即用于电动汽车的功率SiC。这实际上是一个非常好的商机，同时也是技术挑战。当埃隆·马斯克决定从IGBT转向SiC功率模块时，改变了整个行业，而该行业仍在努力生产足够的衬底和逆变器设备，以满足中国和美国等地电动汽车行业的强劲需求。我知道这不仅仅是复苏的问题。

所以，某些领域仍然短缺，如SiC功率器件，某些领域的复苏速度不同，比如晶圆厂和逻辑芯片。你只需要看看英特尔的新闻就能知道，但我们可以看到，在一些领域DRAM和3D NAND业务本应会稍许的回升，因为目前价格还在触底，所以未来将会回升，但这是在生产已经大幅削减的情况下，所以这些工厂仍然面临困境。这都是行业中更多细微差别和细节的信息。

问：那么当前行业复苏的驱动力是什么？

答：是的，我提到了其中一些因素，但从逻辑和晶圆厂的角度来看，人工智能无疑是一个巨大的推动力。比如英伟达是价值1万亿美元的公司，而且正受益于人工智能。毫无疑问，AMD和英特尔正在竞争英伟达的准垄断地位。但事实上，人工智能在各个国家的应用都非常广泛。但说实话，中国和美国是全球人工智能的两个主要国家。对于这些芯片的需求是巨大的。每个人喜欢ChatGPT，但他们不明白的是，ChatGPT每天需要消耗价值75万美元的芯片。

所以提供这些资源的能力很重要，每天需要超过75万美元的资源。所有这些需求不会凭空出现，人工智能无疑是需求的来源之一。我提到了汽车，不仅仅是因为电动汽车的兴起，而且整个汽车行业正在从相对保守的车载微电子技术转向非常激进的技术。这和到底是汽油车、混合动力车还是电动车无关，所有围着汽车转的东西，导航、泊车、安全和通信等方面都在快速增长，这对该行业来说是一个非常好的推动力。

最后我想提到的最后一个是可再生能源。事实上，中国和美国在这方面处于领先地位。当我们正努力发展可再生能源时，从化石燃料转向可再生能源意味着需要大量的半导体和功率半导体来提供各种开关和转换、逆变功能。

你知道我又要提我的特斯拉了，但我也非常自豪地说，加州如今有超过2/3的发电量来自非化石燃料，例如太阳能等。我认为，在人工智能、汽车技术突破和可再生能源方面，我们已经有了足够的动力来推动整个行业的发展。

问：请问各个细分市场供求周期不匹配的现象在半导体行业历史上是否常见？

答：老实说，这是很常见的情况，行业从来都不会停滞不前。总是有进步和变革。如果是一个技术突破较少的行业，你会看到一些稳定性。但每当出现突破，比如SiC功率模块或人工智能，市场就会突然转变。举个例子，汽车制造过去通常采用瑞萨（Renesas）、英飞凌（Infineon）和意法半导体（STMicroelectronics）等公司的落后制程的技术。

话说回来，如果你了解最新的车辆，顺便说一句我喜欢我的特斯拉，但像特斯拉这样的最新车辆实际上使用先进制程工艺。它们使用的微芯片只有5nm 制程。它们正在推动人工智能技术的发展。所以，从65纳米到28纳米，再到5纳米的跨越是巨大的。因此，当你看到供应过剩或供应短缺时，其中一个原因是这个行业本身发生了变化，在功率器件也是如此。过去每个人都使用IGBT（绝缘栅双极性晶体管）。IGBT是适用于高功率应用的高功率双极晶体管，不仅仅局限在汽车和像丰田普锐斯（Toyota Prius）的混合动力车型上。

但现在我们正转向SiC，对SiC器件和衬底的需求突然暴增。我指的是行业正在从IGBT过渡到SiC，但实际上并没有准备好支持切换到SiC的基础和产能，正如我所见，特别是在SiC衬底方面，你看看投资公司的数量和对SiC衬底的投资规模，就能了解到SiC衬底的短缺情况有多么严重。

我同意你的观点。行业总是经历一些变化，从供不应求到供过于求再回到供不应求，但我认为人工智能和SiC等技术突破正在改变这个行业。这就是半导体行业的故事。它每隔几年都会发生重塑。

问：上一个周期的行业状况是否也会影响到下一个周期的行业状况？

答：毫无疑问，新冠疫情使得移动计算的需求远远超出了行业的准备。全球都感受到了汽车行业和消费者以及高性能计算领域的短缺问题。你的观点是，现在存在着严重的短缺，所以我们可能会看到严重的供应过剩。很可能会发生这种情况。我认为，大家都想知道下一次的供求失衡什么时候会发生。在我看来，那将是在2025年左右，距离现在并不遥远。这意味着我们现在想要看到的补贴和投资最终将导致供应过剩。但更重要的是，我认为供应过剩通常是因为私营或公共公司投入过多，因为他们很难预测未来。

现在，我们正在谈论试图区域化或本地化生产，所有地区都想这样做。同样，美国想增加制造业，中国也想增加制造业。印度、日本、欧洲都希望增加制造业。这不仅会推动供应过剩，还会推动成本上升。毫无疑问，一旦开始本地化，虽然这很美好，但这意味着你需要开始削减单位生产的数量，因为全球生产厂变多了，这将推动成本上升。所以，我对此并不乐观。更严重的短缺正在改变下一个周期。我想说的是，下一个周期将严重偏离正常周期。我并不是说情况会变得更好或者更坏，SVRI只是试图帮助客户了解下一个周期将是什么样子。

以及机遇将在哪里？我知道肯定会面临挑战。但老实说，将会有非常强大的机遇，因为后疫情时期对芯片的需求没有回到疫情前的水平，在全球范围内，所有领域的芯片消耗量都有所增加——无论是消费者、计算机、人工智能还是汽车等领域，疫情提高了产业的复合年均增长率，作为一个半导体爱好者，我对此感到非常高兴。

问：AI 芯片的发展速度是否无法赶上 AI 需求？

答：我不这么认为。顺便说一句，我理解你的迫不及待，目前我们本应该有更多的人工智能。我的意思是，我的女儿们喜欢ChatGPT，希望能够随时使用它。但是，我不认为需要几年时间。我认为挑战并不在我们需要建设更多产能，而是在于人工智能技术整体仍在发展中。例如，今天英伟达人工智能芯片也就是GPU的短缺与晶圆厂无关。短缺这不是因为台积电无法生产足够的5纳米芯片，而是这款芯片需要先进的封装，比如CoWoS类型，这才是目前的瓶颈。这是第一点。第二点是，GPU图形处理单元并不是人工智能的终极解决方案，它可能是我们今天所拥有的最接近的解决方案，但你可能也知道有公司正在研发新的设计。

上次我在上海的时候，实际上我遇到了一家初创企业，他们正在开发新的芯片和人工智能概念，如果这一切实现，意味着我们所建立的基于GPU的AI技术在未来不再是需求，我们需要进行转型。所以，我想告诉你的是，我理解你急切的心情，因为我自己也一样迫切地期待人工智能的发展，我也希望随时随地都能用上人工智能。但我认为这些技术还没稳定下来。

我认为尽管英伟达如今赚了很多钱，但我对中国的一些初创企业、AMD和英特尔会在一段时间内推动人工智能迎头赶上并发展壮大非常有信心。我不认为这需要几年时间，可能六个月就够了。最后，所有人都在谈论，例如ChatGPT在性能上尚不稳定的事实。大家都知道ChatGPT存在一些波动需要修复。它是一个不断变化的目标，但不是几年的时间，我认为只需要几个月。

问：目前，根据 SIA 的报告，美国半导体行业整面临严重的工人短缺，不少学术机构加入机构网络来帮助培养半导体人才，您对这个培训渠道有何看法？

答：有两点。第一，是的，毫无疑问，让我们举两个例子：以韩国和美国为例，过去美国的芯片制造只是外包到亚洲，所以并不需要那么多熟练的工程师。需要注意的是，美国仍然拥有大量熟练的工程师，但他们主要面向芯片设计。大部分无晶圆厂公司都是美国企业。这是可以接受的。现在当我们把芯片制造迁回美国时，我们需要找到这些熟练的工程师。这就是美国的例子，但是以韩国为例，他们一直在半导体领域非常强大。让我给你举个例子。最近对技术的禁运意味着SK海力士和三星在中国的投资面临风险。这意味着三星和SK海力士需要撤回并在韩国进一步扩大制造能力，你猜发生了什么？这个计划没考虑人力资源方面。

所以，我记得你提到了美国，但是如果你看看韩国政府，他们非常积极去培养人才。举例来说，如果你看看三星，他们为培养未来的工人而开展的各种活动是巨大的。回到你的问题。顺便说一下，我们的一个客户要求对此进行基准测试，所以我们为他们创建了一份关于中国、韩国、日本、美国和欧洲如何提升熟练技术人员水平的基准报告。我只想提一下在世界各地看到的一些非常好的创意，你提到了大学，这在任何地方都是如此。但是在美国，这是由英特尔和台积电共同推动的，他们现在提供半导体的速成课程，从事真空技术的人才他们不会无师自通制造半导体，但通过速成课程，可以让他们成为半导体工程师。

第二个例子是，这个在三星非常有名。三星已经请回了所有退休的工程师来帮助培训，因为他们没有足够的人手。顺便说一下，我自己也不算很年轻，但我认为吸引过去5到10年退休的工程师成为年轻人的教师、导师、培训师是一个非常棒的想法。是的，我们的确将面临人力短缺问题，但现在也已经有很多出色的解决思路了。不过，我认为这些想法需要进行定制和本地化。例如，中国不会完全照搬韩国的做法，日本也不会完全照搬。但我们的一位日本客户正在考虑根据我们的基准报告，建立某种形式的培训中心，并吸引人才来帮助培训其他人。

所以是的，毫无疑问，我们需要在这个行业中拥有更多的人才。人才数量并不是真正的问题，问题在于人才梯队必须本地化，并且在所有关键地区附近，包括印度、中国和韩国，所有这些地区都将需要更多的人才。其中一些地区，比如韩国，已经拥有非常强大的基础，但是例如印度的基础还不够强大。老实说，在日本，半导体产业已经被认为是过去的东西了。它不再被视为世界上最酷的技术。我是说，在日本，20年前半导体很酷，但是现在年轻工程师不想在半导体领域工作。

我认为准备劳动力，寻找工程师，激励年轻工程师加入半导体行业是一个非常重要的课题。毫无疑问。

问：这种模式，对于中国有什么启示？

答：显然，正如我上面提到的，中国在某种程度上处于一定的位置，所以我给出了一个例子。印度处于前半导体制造时期，而日本则是后半导体制造时期。中国更多地参考韩国的做法。因此，从产业增长周期来看，我认为中国更像韩国，意味着强大的行业快速发展。因此，如果中国想学习如何培养熟练工人队伍，我强烈推荐关注三星正在做的事情。

我前面提到过三星正在招募退休人员参与培训，但韩国政府也发起了大规模的培训计划。他们建立了一个培训中心，和所有学校合作并得到了所有关键公司（如三星、SK海力士等）的支持。我记得他们去年9月签署了备忘录，并计划在今年9月完成第一批400名工程师的培养。因此，对于中国来说，我强烈建议借鉴三星和韩国政府的做法。我知道中国的产业经济规模与韩国不同，但从寻找最佳实践的角度来看，我肯定会推荐这样做。

问：对于日本发展 GAA 的看法？

答：我先打个岔，我曾在NEC（日本电气）工作了10年，当时NEC是全球第二大半导体公司，我在日本公司任职的十年里，我对日本的专业知识非常赞赏，但是从芯片制造角度来看，其半导体行业在某些方面表现出色。其中之一就是CMOS传感器。我知道你提到了索尼，但索尼并不关心逻辑芯片，他们专注于图像传感器，在这个领域他们是世界第一。他们成功击败了三星和意法半导体。所以，从技术角度来看，我认为索尼在图像传感器和铠侠在3D NAND闪存方面处于领先地位。顺便我想纠正一下，大家都认为铠侠是全球第二、第三或第四大闪存制造商，实际上他们的制造规模和三星一样大，人们很少意识到，西部数据和铠侠在半导体制造上实际上是同一家公司。在服务器市场上，他们是全球第一。

如果我们看看日本，在两个领域他们表现得非常出色，即图像传感器和3D NAND。我知道这不是主流产品，也不是主流的代工厂，但我想提醒大家，在使用苹果iPhone时，实际上内部搭载了很多索尼和铠侠的日本芯片。这是第一点。

第二点是，是的，没错。像其他国家一样，日本也希望在逻辑芯片制造方面重返一些领域。他们有两个努力的方向。一个是与台积电合作的28纳米工艺。这里需要强调一下，28纳米工艺处于技术节点中的甜点位置。实际上，这种工艺是成本最低廉的制造方式。另一个是与IBM合作的2纳米工艺。我也曾在台积电工作过，对该公司非常了解。我参与了他们在新加坡的一座晶圆厂建设，因此我非常喜欢代工的工作。

这是否真的是日本正在重塑芯片制造业？我不这么认为。我认为他们是吸引台积电来满足日本对传统制程技术的需求。我知道，传统制程技术，像28纳米芯片仍然是先进的，但这正是台积电擅长的。所以这是第一点。第二点是与IBM合作的2纳米芯片，的确，这比28纳米芯片更不寻常。如果你看看最近的情况，虽然我不能透露我们的客户，但我们与其中一些活动关系非常密切。

我指的是SVRI。但在这种情况下，最初规模相对较小的投资正在变成非常大的投资。对此我印象非常深刻。我的很多日本朋友将加入到这个2纳米的先进节点。我只想提一点，毫无疑问，日本的2纳米芯片项目想要与台积电竞争，但他们正在努力参与其中。因为说实话，台积电是块硬骨头，要与台积电竞争并非易事，但他们至少有一个策略规划，搞清楚哪些领域台积电可能还没触及到。这仍然是一个困难的领域。

但至少我想提到的是，如果你了解他们正在做的事情，你会发现他们正在深思熟虑自己能做什么，有哪些尖端的领域台积电能做，又有哪些不能做。总之不要和台积电硬碰硬，而是寻找一个台积电尚未触及的赛道。最近的一个例子就是人工智能。我不确定这是否会成功，但我对此表示赞赏。我下个月将回到日本，我相信我会以某种方式参与其中。

问：在熊本设厂是否意味着索尼在图像传感器上的限制有所缓解？

答：毫无疑问，我认为图像传感器胜负仍然未知，三星过去曾经是第一，所以我相信三星会尽其所能夺回索尼的领先地位。现在，索尼能否从台积电那里学到一些东西呢？也许吧。老实说，我不认为熊本的意图是这样的。我的意思是，熊本更多的是在逻辑芯片方面有所作为，而不是关于图像方面的。但同时，由于熊本将投产28纳米芯片，也就是采用浸没式光刻技术，这将为逻辑芯片带来很多浸没式光刻技术的专业知识。这能否转化为对图像传感器的支持呢？

毫无疑问会的，日本没有很多具备浸没式光刻技术的半导体制造工厂，尤其是像铠侠那样专注于3D NAND闪存技术的半导体制造工厂不会有。唯一可能拥有较强浸没式光刻技术的地方可能就是广岛的美光（Micron）了。所以，毫无疑问，熊本将成为浸没式光刻技术的第二个卓越中心。这是否能转化为对索尼的支持呢？我会在下个月回到日本，所以关于这个问题还有很多我可以说，我会及时更新的。

问：苹果只像TSMC购买成熟晶圆？

答：首先，关于代工厂的业务方式有两种。他们要么销售晶圆，要么销售好的裸片。通常来说，如果看台积电的业务，大约三分之一的业务是基于晶圆的，也就是说只要经过电测试合格，无论产量如何，他们都会销售出去。而另外的三分之二则是销售成品裸片，这意味着你只需要为晶圆上的合格芯片付款。现在，关于苹果和台积电的问题，没有人会否认苹果是台积电非常重要的客户，也没有人否认在投产的前六个月——这适用于台积电的所有技术——投产后前6个月的制造都处在良率爬坡。他们需要学习以及怎么制造晶圆。

顺便说一句，他们使用了很多人工智能来实现这一点。所以，他们需要很多英伟达芯片来运行他们的生产。但是他们需要大约6个月的时间来调试他们的3纳米工艺。苹果是第一个客户。所以，为什么台积电会同意进行成品裸片销售而不是晶圆销售？很简单，他们与苹果合作，使得他们的3纳米工艺成为世界上最好的技术。无论由于苹果只为成品裸片付款，给台积电少付了多少钱，台积电也会在将来其他客户转向3纳米工艺时弥补这些损失。所以不要担心，我知道这总是成为重点新闻。顺便说一句，每次台积电开展新技术时，通常会有三到六个月的负面报道，说台积电不行了。然后6个月过后，台积电突然又行了。我觉得对你的问题的答案很简单，一旦他们调试好3纳米工艺，他们就会赶超三星。你还记得吧，三星比台积电早大约三到六个月开始3纳米工艺，一旦他们良率爬上去了，台积电就会重新领先。这很明显的事儿。

问：TSMC是否受到了三星的压力？

答：毫无疑问是的，这不是第一次，也不会是最后一次。其中一个原因是三星不仅仅是一个晶圆代工厂，他们是一个集成设备制造商（IDM），这意味着当他们开始尝试他们的技术时，他们有一个试验田，也就是他们手机产品线的基本盘。而台积电没有这样的条件，台积电只能在芯片上进行试验。然后他们需要像苹果这样的客户来调试技术并推动落地到产品。

所以，这不是第一次，也不会是最后一次。通常情况下，因为三星在启动代工业务之前就开始了IDM业务，所以他们与英特尔一样，他们在量产可能会稍微领先于台积电。这么做有啥好处？到目前为止，好像没有。如果你观察三星与台积电市场份额的历史记录，你会发现三星并没有获得显著的市场份额增长。但老实说，我隔壁就是台积电，我也不是不喜欢台积电，我认为台积电不会成为垄断商。我完全支持英特尔和三星获得更多的市场份额。

问：您认为GAA哪家会更有优势？

答：有两点。第一，三星决定在3纳米技术上采用Gate-All-Around结构，我认为这可能不是最好的选择。我这样说可能不会被太多人支持，但是在一个节点内调整晶体管技术总是非常冒险的。从性能、成本和尺寸几个方面来看，三星的3纳米GAA技术并没有特别突出。这是第一点。

第二点是，就纯粹的晶体管技术而言，没有比英特尔更优秀的了。毫无疑问，在全球范围内，英特尔拥有最出色的设备工程师。问题在于，代工厂不仅仅是卖芯片上的晶体管，而是卖提供给客户的商业服务。这就是挑战所在。大家都在试图比较三星、英特尔和台积电的在技术水平、良率或产能水平的差异。我认为这是错误的观点。你应该关注的是哪家公司向其代工客户提供了最好的客户服务。我给你举个例子，仅此供参考。

我不会透露有关知识产权内容。但是作为台积电的一名晶圆厂工程总监，我经常会在凌晨2点接到来自高通（位于圣地亚哥）的电话，跟我说晶圆生产线上出现了什么问题，比如我生产的X号晶圆在三个小时前停止工作了，他们希望我调查并采取措施。这意味着即使是远在圣地亚哥的高通工程师也认为台积电就像他自己的晶圆厂，而他就像在这个晶圆厂里工作一样。我认为目前三星和英特尔在客户服务方面还没有达到这个水平。所以回到你的问题，谁在GAA方面做得最好，我认为是英特尔。但对于代工的角度来说这是否重要？也不是不重要，但这不一定是获取市场份额最重要的方面。

问：IDM时代是已经结束？

答：我的回答可能是，IDM（集成设备制造商）的时代已经结束，我指的IDM是那些单纯设计、制造并最终销售芯片的公司。从今天来看，最大的公司并不是IDM公司，他们是系统制造商，或者叫服务提供商。所以，如果你看苹果、谷歌或者Facebook，实际上芯片在他们的业务中只是一个小组成部分，虽然它很重要。但这些公司之所以成功，是因为他们不卖系统，而是卖服务。比如说，谷歌就在销售各种服务。为了做到这一点，他们必须相台积电购买芯片。再比如说，现在的英特尔仍在销售微处理器，他们并不是系统制造商。如果你不小心或者没有正确判断市场需求，你可能会积压很多没人愿意购买的库存。我的回答是，我认为 IDM 的时代已经结束了。

现在推动产业增长的是服务产品提供商、系统提供商。这是更高一层的定位，芯片不是重点，电路板也不是重点，重点是向客户提供完整解决方案。苹果绝对是一个非常注重客户使用体验和服务满意度的公司。

展望未来，我认为英特尔的IDM战略是正确的。我支持IDM 2.0，并且我真诚地认为现任CEO正在朝着正确的方向发展，但他仍然背负着旧英特尔的包袱。光靠销售微处理器挣钱的时代已经结束了。我可以想象在未来，英特尔将成为一个强大的晶圆代工厂。但他们仍然需要与台积电竞争，张忠谋（Morris Chang）是发明了晶圆代工业务。

我认为世界上很少有人真正理解晶圆代工的含义。大多数人只关注台积电的纳米制程、资金和良率比较。事实上，当企业希望将自己的系统推向市场时，他们希望能够使唤像台积电这样的供应商，就像他们自己的晶圆厂一样。这就是关键。简而言之，我认为IDM销售芯片的时代已经过去了。

问： 为什么政府今天如此积极参与？

答：这问题简单，就是因为缺芯。没人对在疫情期间的缺芯事件感到满意。各国政府的反应是：“那么我就自己生产。”而且，这不仅仅是芯片的问题。口罩也出现了短缺。在疫情期间，每个国家都决定要自己生产口罩。我理解这一点，虽然建立口罩生产线并不是特别昂贵。

但是现在，如果你看一下补贴水平，我之前提到过，我们已经达到了1万亿美元等级的政府补贴。大部分补贴用于将资源和芯片制造业本地化、区域化到自己国家。顺便声明，我并不对补贴行为进行评判。作为一名顾问，我的角色是帮助人们了解行业发展方向，理解行业的发展趋势。我并不是来评判这是否是正确的抉择。我只是把现象忠实地表述出来，现实是如今这些补贴正在推动产业的区域化趋势。

问：美联储推出的货币政策对政府补贴有何影响？

答：显然，SVRI并未涉及行业的金融方面或补贴来源。只是我想提一下，比方在美国，补贴水平只是一个微不足道的金额。如果你看一下美国的GDP规模，这是数千亿美元的级别。与之相比，补贴金额微不足道。所以，说实话，我对联邦银行按月决定的金额与这笔钱之间的联系并不太担心，你可能也知道，但在美国政府预算中，500亿美元意味着什么呢？不多，我觉得反而很少。顺便说一句，这可能在其他地方不成立。我提到的是美国的5.2万亿美元，是的，可能会略有不同。例如，德国目前正在大量投入补贴。与其国内生产总值相比，补贴的比例比美国要高得多，但我仍然不担心，因为德国拥有非常强大的工业和商业实力。但是，毕竟，不同国家之间可能会有所不同。

问：您对这些政府的干预有何看法？

答：是的，我认为这个问题可以从两个甚至三个方面来看。第一个是补贴。补贴是简单明了的，它可以激励人们进行投资，这没啥问题。我觉得虽然补贴会产生干扰，但这是一种可接受的干扰。但是当我们谈到华为时，我们谈论的是限制和技术控制。这些是对产业施加的限制。毫无疑问，相对于补贴而言，这些要更具破坏性。所以补贴还好。我给你举个例子，在中国，政府大约投入了超过1000亿美元来支持本土半导体设备制造商。这是很好的，意味着补贴有助于他们的发展并提高设备水平。

然而，当政府开始介入并对技术进行监管和控制时，这要分成两部分。首先，这与半导体行业无关，我只是作为一个被动的旁观者。第二点，在我看来毫无疑问。例如ASML最终同意了美国的要求，停止向中国出口某些型号的光刻设备后，这将减缓中国在先进工艺制程方面的一些增长。

这意味着什么？这意味着这些投资将不得不流向其他地方。现在问题是，这会造成延迟吗？很可能会，因此，这意味着原本应该在中国进行的一些14纳米晶圆厂的工作将不得不转移到其他地方。我相信这将导致一些投资的延迟。最近，我参与了在印度的一个项目，我毫不怀疑当某个地区无法投资和发展时，它就会寻求其他地区来接手，这正是今天正在发生的事情。所以回到你的问题：这些干预会有负面影响吗？注意我说的不是补贴，而是干预措施、技术控制、知识产权控制是否会影响行业，导致增长放缓或引发短缺，答案当然是肯定的。现在，我毫不怀疑这些措施并不是出于产业目的，而是出于政治原因。我不评价，我只是看到了这一现象。

顺便说一句，SVRI的角色是解释这些限制的后果，因为这些限制主要是对商业模式的干扰。问题是，是否存在机遇？是存在机遇的，只是这些机遇将会转移、改变。我们在SVRI所做的就是了解这些后果，了解在印度、中国、日本等地发生的事情，以及公司应该如何根据这些限制调整他们的策略。回到你的问题，毫无疑问，对技术出口和知识产权控制的某些限制可能会增加一些短缺，可能会延迟投资，改变投资方向。

问：您能给我们一些关于 SVRI 如何帮助该行业的想法吗？

答：是的，我们刚刚讨论了行业的现状。毫无疑问，这是一个复杂的行业。你提到了图像传感器，我们讨论了人工智能，还有碳化硅，这个行业真的很难懂。我们的角色就是帮助行业读懂行情。那么我们如何提供帮助呢？我们为客户提供市场研究报告。但是正如我之前所说，我们努力超越数据和图表。很多咨询公司会给你提供一个300页的报告，里面充斥着大量的图表和数据，但可能并没有得出一个结论，他们只是给你一堆数字，然后让你自己去决定该做什么。而我们的目标并不是这样。我们真的想与客户合作，达成共识，这意味着进行预测和基准测试。我之前提到过，我们正在对技能和员工培训进行基准测试，我认为这非常重要。

所以，我们有一个日本客户想要了解世界上关于培训新技术人员、新工程师和新操作员的最佳方案，这就是我们的工作方式，就是预测和基准测试。而且在某些情况下，我可以给你举个例子。我们曾经有一个客户，我不能透露客户的名字，但他是一个大型代工厂的供应商。他们问了一个问题，就是我们能做些什么来缩短我们的生产周期时间？因为这些代工厂，你可能知道，但我之所以喜欢在台积电工作，就是因为它以分钟为单位，而不是以天为单位。这意味着在台积电的生产周期时间非常快。

所以，供应商受到了这方面的挑战。我们有一个客户要求我们简单地调查他们的制造现场、销售、市场营销和工程部门，告诉他们能做些什么来加快运营速度，因为代工厂的速度非常快，我们很难满足他们的需求。这就是我们所面对的一切。

毫无疑问，如果你看看增长情况，中国半导体产业的增长速度是惊人的。需要说明的是，它可能不是指14纳米制程，也可能是28纳米、碳化硅甚至可再生能源。没有什么能够阻止中国的半导体产业发展，只是发展路线图和增长驱动因素将随着时间而发生变化并适应。这意味着对于很多公司来说，这是机遇。例如，你可能已经了解了碳化硅，最大的碳化硅衬底制造商是美国公司Wolfspeed。他们正在美国扩建最大的碳化硅衬底生产工厂，毫无疑问，其中一些衬底将销往世界各地。

如果你搜索一下Wolfspeed供货协议，就会发现他们与许多设备制造商和电动汽车制造商签订了协议。所以简单地回答你的问题，当这些干扰因素出现时，我们进行区域化，我们正在努力在这些新的地区拓展业务，以发展新的业务。

问：您能否强调一些关键的行业变化 ，尤其是相对以往半导体行业周期里不一样的部分？

答：我认为你刚才说的有句话非常重要：并不是一切都与先进工艺有关。我之后会回答你的问题，但是我对联华电子（UMC）非常了解。我有很多朋友在那里工作，但你可能还记得，在某个时候，联华电子决定不再与台积电在先进工艺上竞争。

看看现在的他们，他们非常赚钱，发展得非常好，但他们并不专注于先进工艺，而是涉足各种应用领域。回到你的问题上，行业中似乎有哪些关键的变革领域？毫无疑问，极紫外光刻技术（EUV）是其中之一。我知道我刚说过先进工艺并不是一切，但是毫无疑问，极紫外光刻技术（EUV）是一个突破。如今，EUV有五家主要用户，包括英特尔、台积电、三星、海力士和美光。当然，在不久的将来会有更多，但这些公司确实推动了光刻技术发展，这就是第一点。

另一个我想提到的但和纯技术没啥关系的领域是工业4.0。人们对工业4.0并不太熟悉，但它与我之前提到的围绕着系统供应商（的生态系统）密切相关。还记得吗？我说过集成器件制造商（IDM）时代已经结束，而系统供应商正在基于工业4.0进行发展。也许你已经知道，工业4.0基本上是一个传感器网络，它能让你全面了解整个生产线，而不仅仅是芯片或电路板，它涵盖了整个生产过程，直到最终产品。我认为这一点并没有得到广泛的理解，但某些公司已经成功地应用了工业4.0。

如果你看一下英飞凌、高塔、台积电等晶圆代工厂，有很多这样的例子。还有一个我现在非常感兴趣的领域，我自己在技术、行业和基础设施领域切换了好几次赛道。我只是想提一下设备开发。我在设备开发方面度过了大约10到15年的时间，担任过总经理职位。关键在于，我们可能正处在一个转折点上，也许你还记得很久以前美国拥有过类似的模式，特别是在设备制造方面。

然后日本崛起了，在接下来的10到15年里，日本占据了市场份额的三分之一到一半。其中包括尼康（Nikon）、佳能（Canon）、东京电子（Tokyo Electron）等公司。现在您猜怎么着？中国现在正在做同样的事情，中国正在走当年日本走的路。所以，我想象在5年后，（我知道我忽略了阿斯麦）除了购买美国或日本的设备，你还将拥有三家供应商：中国、日本和美国。为什么我认为这是好事呢？因为更多的竞争意味着更多的创新，这将促使半导体行业蓬勃发展，增长更快。不好意思，我浏览了业界中几个重要的领域，但这些可能是我列表中最重要的三个点。

以下是采访原文（英文）：

Q:What is SVRI busy with these days, what keeps you awake at night?

A:A main theme is the industry recovery, you know that during Covid, we've been in a shortage situation ramp. After Covid, we went into a ramp down and oversupply, and we are now finally recovering from that, So the industry is recovering, it’s starting to ramp. But the unusual aspect of this recovery is there's a huge amount of government interventions and subsidies. I think the last time I checked we are at a trillion-dollar subsidies by various governments in the world and that's very unusual, so it's going to be a recovery with a lot of distortions.

Q:You talk about semi industry recovery but what is the industry status?

A:Yes, that's a good question, because actually when we talk about the industry, we tend to merge everything in one big industry status. But that's not correct. Actually, if you look at the foundry, the logic, the DRAM, the 3D and flash, the status of recovery is very different. So, I tend to say that we are recovering, but these segments are various levels. I think definitely that the foundry and maybe the logic is head of the pack. That's usually the case. But there are good reasons for that. And I think DRAM and 3D NAND is barely plateauing at the bottom of price curve. So, some of our customers, by the way, are memory makers, and they are not yet happy with the situation. We know it's going to come back, but it's not exactly where we should be.

And then I mentioned that we are over shortages, but that's not completely correct. I know that there's always somebody telling me we are still in shortage. There are few areas of shortage and I just want to mention one which is the power SiC for EV vehicles. This is a fantastic actually business opportunity, but also technology. When Elon Musk decided to switch over from IGBTs to SiC power modules, e-change the industry, and the industry is still trying to generate enough substrate, enough for inverter devices, for the EV industry, which is ramping strongly in China, as well as in the US, I know there's a little bit more neurons than just recovery.

So, some areas still in shortage, like sic power, some areas are differently in the ramp like foundries and logic. You just need to read the letters from Intel, but some areas DRAM and 3D NAND would like to see a little bit more business coming, you know the pricing still is at the bottom. It's going to come back up, but there's been a lot of cut in production, so these factories are still suffering. So that's a bit more nuance and details on the industry.

Q:So what are the current industry recovery drivers?

A:Yes, I mentioned some of them, but from a logic and foundry point of view, artificial intelligence no doubt is a huge driver. I mean Nvidia by the way the $1 trillion company and Nvidia is benefiting from it. There's no question that AMD and Intel are going after the quasi monopoly of Nvidia. But the fact is that all the artificial intelligence usage in every country. But to be honest, China is and US are the top two artificial intelligence countries in the world. The demands for these chips are huge. Everybody is love ChatGPT what they don't understand is that ChatGPT consumes three quarter million dollars of microchips a day.

So there are capacities. It is above $750,000 a day. All of these needs to come from somewhere. So AI is definitely one. I’ve mentioned automotive not just the fact that they are EVS, electrical vehicles, but the automotive is switching over from a relatively conservative technology in terms of microelectronics on board to very aggressive. It doesn't matter. Actually, if the car is gasoline, hybrid or electrical, all of these support around the car in terms of routing, parking, safety, communications are exploding and this is a very good driver for the industry.

And then the last one I want to mention, I mentioned a little bit as I see, but renewable. The fact that and again, China and the US are really leading the pack there. The fact that we are trying to work on renewable energies, moving from fossil fuel to renewable means that a lot of the semiconductor, power semiconductor needs to provide all of these switching capabilities, inversions, doing this.

You know, I mentioned my Tesla, but I’m also very proud that, you know, California today is over 2/3 non-fossil fuel, electricity production, solar, and all of these. That's, I think, between AI the automotive breakthrough and the renewable energies, I think we have more than enough to drive the industry.

Q:Is the mismatch between the supply and demand cycles of various market segments common in the history of the semiconductor industry?

A:It is common, to be honest, the industry is never a standingstill. There are always progress. If it was an industry with less technological breakthrough, you would see some stability. But every time the breakthrough, like, for example, SiC power modules or artificial intelligence, then suddenly the market shifts, I give you an example, but automotive tended to be productions of lagging type technologies by companies like Renesas, Infineon, STMicroelectronics.

The latest cars, if you know, by the way, I enjoy my Tesla, I love it, but the latest cars like the Tesla is actually leading edge. These are microchips which are in the 5 millimeters. They are pushing the envelope with AI, so the shift from maybe 65nm, 90nm to 28nm, 9nm, to 5 nanometer is a big shift. And so when you see the oversupply, undersupply shortage, one of the reason is the industry is changing same way for the power. Everybody used to use IGBTs. IGBTs are high powered bipolar transistors for high power, just not only automotive and hybrid cars like Toyota Prius.

But now that we've moved to SiC suddenly, the demand for SiC devices and substrates is humongous. I mean in the industry is switching over from IGBT to SiC without actually having established a good ground and a good capacity for all these, as I see, and especially the substrate where you look at the number of companies investing and the amount of investment in SiC substrate, it tells you how bad the shortages in SiC substrate.

I agree with you. There's always a bit of change and from supply to oversupply undersupply, but I think breakthroughs in, I mentioned them, but AI SiC are just changing the industry. And this is the story of the semiconductor industry. It recreates itself every few years.

Q:Will the industry status in the previous cycle also affect the industry status in the next cycle?

A:There's no question that Covid drove the demand for a mobile computing way beyond what the industry was ready for. And the shortage, which was both automotive, but consumer, high power consumer computing, was felt all over the world. Your point was, there was a severe shortage, so we are gonna see a severe oversupply. Probably, yes. I think that I can expect, you know, everybody wants to know when is going to be the next crash. In my mind, it's 2025, not that far from now. It means that the subsidies we want to see now, the investment we see will end up with some oversupply. But more than that, I think oversupply is usually when private or public companies invest too much, because they have difficulties to forecast.

Now, we are talking about trying to regionalize or localize the production and all regions want to do that. So, the same way, the US wants to do more manufacturing, China wants to do more manufacturing. India wants to do more manufacturing, Japan, Europe. That will not only be driving oversupply, but that will drive cost. There is no question in my mind that once you start to localize, it's nice, but that means that you need to start to scale down the units of production, because you multiply them all over the world. And that is going to be driving costs. So, I’m not very happy with that to a factor. Deeper shortage is changing the next cycle. And I would say that the next cycle is going to be severely distorted. I’m not saying it's bad or good, SVRI is just trying to help customers understand what will the next cycle will look like.

And what will be the opportunities? I know there will be challenges. But to be honest, they are going to be very strong opportunities, because the demand from post-Covid did not go back to a pre-Covid level, we stepped up our consumption of microchips in the world everywhere, consumer, computing, AI, automotive, Covid boosted , the CAGRs of the industry, and you know being a semiconductor fan, I’m very happy about that.

Q:So we need to find the foundaries and led them to put to produce more AI chips. So maybe the short supply the cycle maybe will continue a few years. Do you believe or something like that?

A:No, I don't believe so. Although, by the way, I’m very comfortable with your impatience, we should have more AI today. I mean, you know my daughters of AI ChatGPT, they just want to be able to use it anytime whenever they want. But so, no, I don't think years. I think the challenge is not so much that we need to build capacity. The challenge is that AI technology is still evolving. For example, today, the shortage of Nvidia AI chips, GPUs is not related to foundry. It's not the fact that TSMC cannot produce enough 5 nanometers microchip. It’s the fact that this chip needs to be advanced package, in advanced packages like CoWoS type. That is today’s bottleneck. That's number one. Number two is that actually, a GPU, Graphic Processing Unit is not the ultimate solution for AI, it's probably the closest we have today, but probably that there are companies working on new design.

And last time I was in Shanghai, actually I met a startup, working on the developing new chips, new concept for AI, if that happens, it means that whatever we've built on GPUs will not be the demand of tomorrow, we have to ramp. So, my point to you is, I love your impatience, I mean I’m also impatient. I want AI all the time, but I think that these technologies not yet stabilized.

I think that although Nvidia is doing a lot of money today, I’m very comfortable with the idea that some of the Chinese startup, AMD, Intel, will definitely help AI match you over time and ramp. I don't think it's a matter of years. But it's probably a matter of 6 months. Last comment is everybody talks about that, the fact that, for example, ChatGPT is not yet stable in terms of performance. Everybody knows that ChatGPT as fluctuations that needs to be fixed. It's a moving target, but not years, I think, months.

Q: So the US government and they will quite a few of economic agencies join the network to institutions, to help build semiconductor workforce. What's your opinion on this phenomenon?

A:Two things. Number one, there's no question，let's take two examples, Korea and USA, let's say for USA in the past, manufacturing was just outsourced to Asia. And so there was no need of so many skilled engineers. By the way, we have a lot of skilled engineers, but these are designers. The fabless companies are massively Americans. That's okay. Now, when we pull back the manufacturing, we need to find these skilled engineers. That's the example of USA, but for example, Korea, which was always has been very strong in semiconductor. Let me give you an example. The recent embargo on technology means that the SK Hynix and Samsung investment in China are at risk. What does it mean? It means that Samsung and SK Hynix need to pull back and ramp up further in Korea the manufacturing, guess what? They did not plan that from a manpower point of view.

So if you look at you mentioned, I think the USA, but if you look at the Korean government, they are extremely active. So, if you look at, for example, Samsung, the amount of activities trying to find and prepare train the workers of the future is massive. Going back to your question. By the way, one of our customers asked to benchmark that, so we actually created for them a benchmark report of all of the activities in China, in Korea, in Japan, in USA, in Europe, of how to ramp up the skilled workforce. I just want to mention few, I think, very nice ideas I’ve seen in the world, you mentioned about universities that's true everywhere, by the way. But in USA, this was driven both by Intel and TSMC, they are now spatial crash courses in semiconductor to people working on vacuum technology, don't to automatically know semiconductor, but with a crash course, you can convert a vacuum engineer into a semiconductor engineer.

One example, second example is, and this is very famous at Samsung. Samsung has asked all their retired engineers to come back and help train, because they don't have the people. And I’m in favor, by the way, I’m not very, very young, but I think that pulling in the retired engineers of the past 5, 10 years to become teachers, mentors, trainers of younger people. That's some fantastic ideas. So, yes, we're going to have shortages of manpower. There are a lot of excellent ideas on the table. Now, I think these ideas have to be customized and localized. So, for example, China will not do exactly what Korea is doing. Japan will not do exactly. But one of our Japanese customers is considering based on our benchmark report, creating some form of training centers and pulling in people to help train these people.

So, yes, no questions. We're going to need more people in the industry. More people is not really the issue, the issue is that they have to be localized and they have to be in nearly all key regions, India, China, Korea, all of these regions are going to need a lot more and some of these regions, Korea is a very strong base, but for example, India has not yet a very strong base. And Japan, to be honest, in Japan, semiconductor is seen as the past. This is not seen as the coolest technology in the world. I mean semiconductor in Japan was cool 20 years ago, today the young engineers don't want to work in semiconductor.

I think preparing the workforce, finding the engineers, motivating the young engineers to get into the semiconductor a business. No question.

Q:What China can learn from this kind of cooperation?

A:Obviously, as I mentioned, China is in a certain position, so I gave you an example. India is pre semiconductor, and Japan is post semiconductor. China refers more to Korea. So, if you look at the cycle of growth in the industry, I think that China is more along the lines of what Korea is doing, which means strong industry growing fast. That's so, I would say that if China wants to learn about preparing the work skilled force, I would definitely recommend looking at what Samsung is doing.

And again, I’ve mentioned the fact that Samsung is pulling in the retirees to do that, but the Korean government launched a massive effort into training. They created a center for training across all universities with support of all the key companies. That means Samsung, SK Hynix, et cetera. I think they signed the memorandum in September last year, and they plan to have their first batch of 400 engineers coming out in September. So for China, I would recommend definitely to look at what is done around Samsung and the Korean government. I know the size of China industry economy is different than Korea, but from finding best known methods, I would definitely recommend that.

Q:They were more focused on with the cooperation with IBM, they were more focused on the GAA so what do you think about that?

A:I want to jump, by the way, I worked for NEC, Nippon Denki shop for 10 years, when NEC was number two semiconductor company, I spent about 10 years working for this Japanese company, and I appreciate the Japanese know how, but the semiconductor industry in terms of manufacturing is actually very good at certain things. One is CMOS sensor. I know you mentioned Sony, but Sony, it doesn't really care about logic. They care about image sensor and they are number one in the world. They managed to defeat Samsung, ST Microelectronics. So, if you look at terms of technologies, I think Sony intimate sense of Kioxia in 3D NAND flash. By the way, I want to correct something. Everybody thinks that Kioxia is a second, third, or 4th in the world. It's actually manufacturing as big as Samsung, people don't realize that, you know, Western Digital and Kioxia are actually the same company when it comes to manufacturing. And they are number one in the world in surf or so.

If you look at Japan, they are two areas where they are really good, image sensor and 3D NAND. I know it's not mainstream object, it's not mainstream foundry, but I just want to remind that if when you use these Apple, iPhone, actually, there's a lot of Sony and Kioxia Japanese microchips inside. So that's number one.

Number two is, yes, it's correct. As everybody else, Japan wants to run back some of the manufacturing in logic, and you are correct. There are two areas of effort. One is 28 nanometers with TSMC and by the way, I just want to remind everybody that 28 nanometer is a sweet spot in the technology node. There's actually the technology which is the cheapest in terms of manufacturing. The second one is 2 nanometers work with IBM, a few things, one is, by the way, I work for TSMC too, so I know that company very well. I build one of their fab in Singapore, so I enjoyed foundry work.

So, one thing, TSMC is working with a number of companies amongst Sony and they are going to be starting ramping production of 20 nanometers in Japan. I don't want to go and discuss government subsidies interventions in that fact. But the fact is simple. TSMC is opening a manufacturing site in Japan, because Japan is a big market for TSMC, when I was building that fab, actually in Singapore, a lot of our customers were Japanese.

Is that really Japan recreating itself? I don't think so. I think that it is pulling in TSMC to supply the needs in Japan of standard technologies. I know, standard technology, 28 nanometers is still advanced, but it's something that TSMC can do very well. So that's number one. Number two is 2 nanometers with IBM, that is a little bit more unusual than the 28 nanometer. And if you look at the latest, and again, I can't disclose our customers, but we are very close to some of these activities.

I mean as we are right. But in this case, what started as a relatively small investment is now turning into a very large investment. I’m actually impressed. I mean a lot of my Japanese friends are going to join that center of excellence at 2 nanometers. Just one comment I want to make is that there's no question that the 2-nanometer effort in Japan wants to compete with TSMC, but they are trying to be part about it. Because to be honest, TSMC is not to crack. It's not easy to compete with TSMC, but they at least have a strategy planning effort to define gaps where they can provide what TSMC maybe will not be able to provide. It's still a difficult area.

But at least I just want to mention that if you look at what they are doing, they are doing a lot of thinking about what can be done and work at the edge of what TSMC can or cannot do. Just don't go ahead with TSMC but work on some of the gaps of what TSMC does not offer. One of the latest is artificial intelligence. I’m not sure if that will stick, but I’m actually appreciate is of that. I'll be back in Japan next month, and I’m sure I'll be involved one way or another.

Q:Maybe that's why they need the co cooperate with the TSMC to build the JSM, because the small things, small pixel size, markets. Nowadays, Samsung, maybe it is. It's the count biggest player in the small pixel size. Market says.</